Overcoming Domain Shift in Violence Detection with Contrastive Consistency Learning

Abstract

Automated violence detection in video surveillance is critical for public safety; however, existing methods frequently suffer notable performance degradation across diverse real-world scenarios due to domain shift. Substantial distributional discrepancies between source training data and target environments severely hinder model generalization, limiting practical deployment. To overcome this, we propose CoMT-VD, a new contrastive Mean Teacher-based violence detection model, engineered for enhanced adaptability in unseen target domains. CoMT-VD innovatively integrates a Mean Teacher architecture to adequately leverage unlabeled target domain data, fostering stable, domain-invariant feature representations by enforcing consistency regularization between student and teacher networks, crucial for bridging the domain gap. Furthermore, to mitigate supervisory noise from pseudo-labels and refine the feature space, CoMT-VD incorporates a dual-strategy contrastive learning module. DCL systematically refines features through intra-sample consistency, minimizing latent space distances for compact representations, and inter-sample consistency, maximizing feature dissimilarity across distinct categories to sharpen decision boundaries. This dual regularization purifies the learned feature space, boosting discriminativeness while mitigating noisy pseudo-labels. Broad evaluations on five benchmark datasets unequivocally demonstrate that CoMT-VD achieves the superior generalization performance (in the four integrated scenarios from five benchmark datasets, the improvements were 5.0∼12.0%, 6.0∼12.5%, 5.0∼11.2%, 5.0∼11.2%, and 6.3∼12.3%, respectively), marking a notable advancement towards robust and reliable real-world violence detection systems.

1. Introduction

Violence detection (VD), a subtask of human action recognition, addresses a research area of substantial practical importance [1,2,3,4,5]. The development of robust violence detection systems is paramount for curbing the proliferation of harmful content on digital platforms, enabling the proactive monitoring of potential violent incidents in real-world environments, and facilitating preventive measures against such behaviors. Collectively, these applications notably contribute to safeguarding public safety and maintaining the integrity of both online and physical spaces.

Driven by notable advancements in artificial intelligence, methodologies for violent detection have undergone a paradigm shift, transitioning from traditional manual feature engineering approaches [1,2] to sophisticated deep learning-based architectures. Current research largely employs three main model categories: (a) hybrid frameworks combining two-dimensional convolutional neural networks (2D CNNs) with temporal modeling modules [6,7,8] to capture both spatial and sequential dynamics; (b) end-to-end spatiotemporal representation learning via three-dimensional convolutional neural networks (3D CNNs) [3,4,5]; and (c) more recently, attention-based architectures, particularly Transformer models [9,10], which excel at modeling long-range dependencies in video data. While these contemporary models often demonstrate strong performance in controlled experimental settings, their efficacy in practical, real-world deployments is frequently hindered by notable challenges, predominantly performance degradation when encountering unseen or dynamically changing environments.

Concretely, the reliable real-world deployment of these models is principally impeded by the domain shift [11] phenomenon. This challenge stems from a fundamental distributional discrepancy between the data used for training and the data encountered during actual deployment [11,12,13,14,15,16,17]. In violence detection, this manifests acutely due to the vast, unpredictable variability inherent in real-world scenarios, which datasets, by their nature, can only partially represent. For example, authentic violent incidents in the real world frequently occur under dynamically complex visual conditions such as occlusions (Figure 1a), where key elements may be obscured, or in low-illumination environments (Figure 1b), where visual clarity is compromised. These characteristics of target domain often differ qualitatively and in their range of variation from the training data subsets, which might include specific scenarios like actions against a pure-color background (Figure 1c) or with high motion blur (Figure 1d). Models trained on limited violence datasets may encounter unknown environmental changes when applied to real-world scenarios. This situation, where the data distribution used during training is inconsistent with the data encountered in practical applications, is a manifestation of domain shift [18].

Such divergence is practically unavoidable, as real-world scenarios are inherently open-ended, encompassing countless variations in environmental factors, camera perspectives, and scene compositions that finite datasets cannot exhaustively capture [12,17,19]. This inherent gap between training and real-world operational data critically exacerbates domain shift, profoundly undermining model generalization and leading to sharp performance declines in practical applications. This degradation carries severe security implications: systems may generate false alarms under benign environmental variations or, more critically, fail to detect genuine threats, potentially enabling exploitation by malicious actors. Addressing this pervasive domain shift is therefore paramount for developing reliable and robust VD technologies.

To mitigate this pervasive domain shift, we introduce CoMT-VD, a new domain adaptation framework specifically designed for violence detection. Our primary objective is to adequately transfer and adapt knowledge from labeled source domains to unlabeled target domains [17,20,21], thereby notably enhancing generalization ability. CoMT-VD strategically leverages the Mean Teacher (MT) architecture [12,13], comprising the student ${\theta }_{\mathcal{S}}$ network and teacher one ${\theta }_{\mathcal{T}}$ with identical structures, where ${\theta }_{\mathcal{T}}$ is updated via Exponential Moving Average (EMA) of ${\theta }_{\mathcal{S}}$. This design choice is fundamental because the EMA mechanism empowers the teacher network to generate highly stable and reliable pseudo-labels for sufficient unlabeled target data. These stable pseudo-labels, in conjunction with a consistency loss, are then critically utilized to guide ${\theta }_{\mathcal{S}}$ training, compelling it to align its predictions with ${\theta }_{\mathcal{T}}$ and thereby facilitating essential adaptation to diverse domain variations. While data augmentation [16,20,22] further supports this adaptation by simulating real-world visual diversity, the inherent noise stemming from imperfect pseudo-labels generated by the MT setup necessitates a more robust feature refinement strategy. To address this crucial need and purify the learned feature space [23,24], we propose and integrate a dual-strategy contrastive learning (DCL) module. DCL meticulously refines feature representations through two complementary objectives. First, intra-sample consistency compels ${\theta }_{\mathcal{S}}$ to produce highly similar features for differently augmented views of the same input. This ensures that learned features are robust to superficial visual changes and capture the essential, invariant characteristics of violence. Second, inter-sample consistency enhances semantic coherence. It actively maps features of samples with similar violent patterns closer together while maximizing distance for dissimilar categories, thereby sharpening decision boundaries. This synergistic integrated approach allows for the Mean Teacher framework to adequately drive domain adaptation [12,13,25], simultaneously ensuring that DCL purifies the learned feature space by enforcing robustness to nuisance visual variations and notably enhancing categorical discriminability. This process adequately mitigates the detrimental impact of potentially noisy pseudo-labels and ensures that the student network ${\theta }_{\mathcal{S}}$ learns highly discriminative yet domain-invariant features, ultimately leading to improved generalization in violence detection across shifting domains.

The main contributions of this paper are summarized as follows:

• We present a pioneering investigation into the pervasive domain shift challenges in violence detection. To adequately address this, we propose CoMT-VD, a new contrastive Mean Teacher model specifically designed to enhance model adaptability and performance across diverse target domain distributions.
• We introduce a new dual-strategy contrastive learning (DCL) module that integrates two distinct positive-negative pair matching strategies to compute complementary consistency losses. This promotes the learning of more effective discriminative features for violence detection under challenging domain shift conditions.
• We conduct comprehensive evaluations of detecting violence under challenging domain shift scenarios, which unequivocally demonstrate consistent and notable performance improvements when CoMT-VD is integrated with various baseline models.

In the following sections, we will introduce current research on violence detection in the Related Work section, along with relevant studies involving domain adaptation, contrastive learning, knowledge distillation, and other related methodologies that this paper will address. In the Proposed Method section, we will provide a detailed explanation of the CoMT-VD training strategy we have developed. In the Experiments and Ablation Studies sections, we will demonstrate how the proposed CoMT-VD enhances model performance in dealing with domain shift issues. Finally, in the Conclusions section and Future Work sub-section, we will summarize the contributions of this paper and discuss the limitations of the current research.

2. Related Work

2.1. Violence Detection

Early efforts for recognizing violent behaviors predominantly relied on handcrafted feature engineering [1,2]. The rapid evolution of artificial intelligence has catalyzed a paradigm shift toward deep learning-based approaches. Recent violent detection architectures are primarily categorized into three technical streams: 2DCNN-based spatio-temporal methods: Extracts spatial features via 2D convolutional networks coupled with temporal modeling modules, e.g., LSTM or TSM [6,7,8], for sequential reasoning. 3DCNN-based methods: Directly captures spatiotemporal correlations through 3D convolutions, e.g., C3D, I3D, or X3D [3,4,5]. Transformer-based: As researchers extend Vision Transformer (ViT) architectures to video understanding, transformer-based methods are increasingly being explored for violent detection tasks. We note that 2DCNN+LSTM network excels in long-term temporal modeling, but struggles to capture short-term local continuous temporal variations. The 3DCNN-based network has limitations in detecting long-duration continuous actions. Transformer-based networks demonstrate superior capability in global relationship modeling for video context understanding, although challenges persist in balancing local feature sensitivity and reducing data dependency. Researchers have also explored avenues such as multi-stream networks to fuse richer multimodal information [26,27].

Despite notable advances, a fundamental and persistent challenge continues to hinder contemporary approaches: the domain shift between training and testing distributions. Models trained on specific, often curated datasets—such as cinematic depictions of violence or controlled surveillance footage—frequently suffer severe performance degradation when deployed in real-world settings. These environments typically exhibit substantial variations in lighting, occlusions, camera viewpoints, recording quality, and the subtle, unpredictable nature of violent behavior. This critical generalization gap, particularly evident in cross-domain scenarios, highlights an urgent need for models with greater adaptability. Addressing this challenge lies at the core of our CoMT-VD framework, which is specifically designed to enhance robustness against such distributional shifts and improve cross-domain generalization.

2.2. Knowledge Distillation and Mean Teacher

Knowledge Distillation (KD) presents a paradigm wherein a student model learns to emulate the decision-making processes of a larger teacher model [28]. This technique has found successful application in various domains, including human action recognition [29], where student models have demonstrated notable performance enhancements by leveraging the distilled knowledge from more potent teacher architectures. One of the important evolutions of KD, particularly for semi-supervised and unsupervised learning scenarios, is the Mean Teacher (MT) framework [13]. In the MT paradigm, the teacher model’s parameters are not independently trained, but are instead maintained as an exponential moving average (EMA) of the student model. This “mean” teacher then provides stable and consistent pseudo-labels for unlabeled data, which in turn guide the student’s training process. This self-ensembling mechanism has exhibited distinct advantages, especially in contexts requiring adaptation to discrepancies in target domain data distributions and the enhancement of multi-scenario generalization capabilities [30,31,32]. For instance, Kumar et al. [30] introduces the stable Mean Teacher, incorporating an error recovery module to refine the quality of pseudo-labels, while Xiong et al. [32] utilizes a teacher memory bank for the generation of pseudo-prototypes.

We observe that the systematic application of the Mean Teacher (MT) framework to address the nuanced challenges of violence detection—particularly under notable domain shift—remains underexplored. Unlike other action recognition tasks, violence detection often hinges on subtle visual and temporal cues, and misclassification can lead to far more severe consequences. Our work is the first to pioneer the adaptation of the MT framework specifically for violence detection. By leveraging its well-established stabilization properties, we generate reliable pseudo-labels from unlabeled target domains, which are typically characterized by real-world complexities such as partial occlusions and poor illumination. In doing so, we address a critical gap in the existing literature and advance the robustness of cross-domain violence detection.

2.3. Contrastive Learning (CL)

CL has rapidly emerged as a potent self-supervised learning paradigm, dedicated to acquiring rich and meaningful feature representations from data. It aims to learn representations by maximizing agreement between augmented views of the same sample (positives) and minimizing it with other samples (negatives) [24,33]. This approach has been successfully applied to various tasks, enabling models to learn domain-invariant features, even in the presence of limited or entirely unlabeled data, by compelling the model to focus on consistent, underlying patterns that transcend superficial domain variations [34,35,36,37]. Illustrative examples include work by [34], who utilizes videos played at different speeds as a basis for temporal contrastive learning, and [35], who augment positive pairs with samples from synchronous viewpoints and further refined the process by using classifier probabilities for hard negative mining. Similarly, Lorre et al. [36] employed contrastive predictive coding (CPC) to learn long-term temporal dependencies that are robust to local noise and variations. Currently, methods based on contrastive learning have been widely adopted by scholars for video content analysis. In [38], a novel form of multi-scale and cross-scale contrastive learning was instantiated to capture salient semantics between video moments. Dave et al. [39] proposed, for the first time, a temporal contrastive loss by designing two types of temporal contrastive losses—local-local and global-local—to distinguish the temporal steps of feature maps in input clips, thereby enhancing the temporal diversity of the learned features and accomplishing video content understanding tasks. Building upon Dave et al. [39], Altabrawee and Noor [40] utilized sparse local clips with a local-local loss to eliminate the need for the second TCLR temporal loss (global-local loss). These studies collectively demonstrate the advantages of contrastive learning in video content analysis.

Nevertheless, contrastive learning (CL) approaches typically emphasize instance-level discrimination or pretext tasks, which may not adequately promote the semantic separability required for effective violence detection—especially when integrated with pseudo-labeling from a Mean Teacher framework. To address this limitation, our CoMT-VD introduces a dual-strategy contrastive learning (DCL) module, explicitly designed to synergize with the Mean Teacher paradigm. DCL enforces both augmentation consistency and feature-level coherence under the guidance of pseudo-labels, aiming to refine the feature space and enhance the discriminability of violent patterns in the presence of domain shift.

2.4. Domain Adaptation (DA)

DA encompasses a set of techniques designed to mitigate the distributional disparity, or “gap”, between a labeled source domain and an unlabeled or sparsely labeled target domain. Addressing this gap is of paramount importance for the successful and reliable deployment of machine learning models in real-world applications, where data heterogeneity is the norm rather than the exception [11]. DA methodologies have undergone notable evolution, progressively incorporating principles from adversarial learning, self-supervised learning, augmentation, and generative modeling, all aimed at achieving greater efficiency, robustness, and dynamic adaptability in bridging domain differences [11,14,15,16,17]. Within the specific context of video-based action recognition, various DA methods have been proposed. For example, approaches like ACAN [14] have employed adversarial training objectives to align pixel-wise correlations across domains, while other research [17] has conceptualized unsupervised domain adaptation (UDA) as a problem of learning from noisy labels, strategically selecting small-loss samples from the target domain based on assessments of pseudo-label correctness. Sohn et al. [41] addressed the challenge of recognizing unlabeled faces in videos based on still-image face recognition by employing an image-to-video feature-level domain adaptation method to learn discriminative video-frame representations. Kim et al. [42] utilized multimodal information to align features using video domain adaptation, achieving cross-domain action recognition. Chen et al. [43] proposed a Temporal Attention-aware Adversarial Adaptation Network (TA3N) to accomplish domain alignment. Aich et al. [44] introduced a novel “Zero-shot Cross-domain Video Anomaly Detection (zxVAD)” framework that learns features by understanding how normal event video features differ from those in pseudo-anomaly examples. Such studies demonstrate the potential of DA techniques to improve cross-domain feature alignment and enhance the reliability of pseudo-labels in the face of disparate data distributions.

Many DA techniques, however, can be complex to implement or may not optimally address the subtle, discriminative cues defining violent acts. Our CoMT-VD framework offers an integrated DA strategy. The Mean Teacher inherently facilitates unsupervised adaptation. We bolster this with strategic cross-domain augmentation (simulating realistic variations like occlusion and low-light) and the DCL module (promoting feature alignment and discriminability). This holistic approach, embedded within the learning framework, aims for robust adaptation and superior generalization in violence detection scenarios affected by domain shifts.

3. Proposed Method

This section presents our proposed approach to violence detection, specifically designed to address key challenges such as domain shift. we introduce the Contrastive Mean Teacher Violence Detection (CoMT-VD) framework, detailing its overall architecture and elaborating on the motivation and design of its core components: the teacher-student paradigm, a dual-strategy contrastive learning mechanism, and the Mean Teacher optimization procedure.

3.1. Problem Definition

Let $\mathcal{X}$ denote the input space, representing video segments. Each video segment can be associated with a label from the label space $\mathcal{Y}=\{0,1\}$, where $y=1$ signifies the presence of violence and $y=0$ signifies its absence. The source domain ${\mathcal{D}}_{sou}$, consists of $|{\mathcal{D}}_{sou}|$ samples and is formally represented as ${\left\{({\mathit{x}}_i^{sou},y_i^{sou})\right\}}_{i=1}^{N_{sou}}$, where ${\mathit{x}}_i^{sou}\in \mathcal{X}$ is the i-th video segment from the source domain and $y_i^{sou}\in \mathcal{Y}$ is its corresponding ground-truth label. In addition, we have access to a target domain, denoted as ${\mathcal{D}}_{tar}$, with $|{\mathcal{D}}_{tar}|$ samples, represented as ${\mathcal{D}}_{tar}={\left\{{\mathit{x}}_j^{tar}\right\}}_{j=1}^{N_{tar}}$, where ${\mathit{x}}_j^{tar}\in \mathcal{X}$ is the j-th video segment from the target domain. Crucially, the labels for the target domain data are often unavailable or only sparsely available, posing a notable challenge.

The core problem addressed is that of cross-domain violence detection. We aim to learn a model $f:\mathcal{X}\to \mathcal{Y}$, which generalizes to the target domain, even when there is a domain shift between the source ${\mathcal{D}}_{sou}$ and target distribution ${\mathcal{D}}_{tar}$. The objective is to optimize f to learn domain-invariant features that are also highly discriminative for the violence detection task. Our approach seeks to achieve this by adequately leveraging any available unlabeled target data and by enhancing the model’s capacity to learn from augmented counterparts of the input data, thereby bridging the gap between domains.

3.2. Contrastive Mean Teacher Violence Detection

To achieve this, we propose the Contrastive Mean Teacher Violence Detection (CoMT-VD) model, which synergistically combines semi-supervised learning via the Mean Teacher paradigm with contrastive learning’s robust representation capabilities, further refined using a cross-domain augmentation strategy.

3.2.1. Method Overview

The CoMT-VD model, depicted in Figure 2, leverages a teacher–student framework [12], where both teacher and student models, w.r.t. $f_{\mathcal{T}}$, $f_{\mathcal{S}}$, parameterized by ${\theta }_{\mathcal{T}}$ and ${\theta }_{\mathcal{S}}$, shares the same network architecture. One critical aspect of our CoMT-VD is the use of cross-domain augmentation: the student model is fed data (from source and target domains) processed by the strong augmenter ${\mathcal{A}}_{strong}$, while the teacher model receives data processed by the weak augmenter ${\mathcal{A}}_{weak}$. As shown in Figure 2, labeled source domain ${\mathcal{D}}_{sou}$ data undergoes the strong augmenter ${\mathcal{A}}_{strong}$ to ${\left\{{\overline{\mathit{x}}}_k^{sou}\right\}}_{k=1}^K$ and is fed to the student network ${\theta }_{\mathcal{S}}$ for supervised learning via ${\mathcal{L}}_{\sup }$. Unlabeled target domain ${\mathcal{D}}_{tar}$ data is processed through the weak augmenter ${\mathcal{A}}_{weak}$ to ${\left\{{\tilde{\mathit{x}}}_k^{tar}\right\}}_{k=1}^K$ for the teacher ${\theta }_{\mathcal{T}}$ and the strong augmenter to ${\left\{{\tilde{\mathit{x}}}_k^{tar}\right\}}_{k=1}^K$ for the student. The teacher, updated via exponential moving average (EMA), provides pseudo-labels from ${\left\{{\tilde{\mathit{x}}}_k^{tar}\right\}}_{k=1}^K$ to guide ${\theta }_{\mathcal{S}}$ through a self-supervised consistency loss ${\mathcal{L}}_{\mathrm{self}}$. The strong-augmented target samples ${\left\{{\overline{\mathit{x}}}_k^{tar}\right\}}_{k=1}^K$ are also input to ${\theta }_{\mathcal{S}}$ to extract features for the dual-strategy contrastive learning (DCL) module, which applies a contrastive loss ${\mathcal{L}}_{\mathrm{con}}$ to learn invariant representations. The framework is optimized with a composite loss with ${\mathcal{L}}_{\sup }$, ${\mathcal{L}}_{\mathrm{self}}$, and ${\mathcal{L}}_{\mathrm{con}}$, enabling effective cross-domain violence detection.

The student model $f_{\mathcal{S}}$ learns by minimizing a supervised loss on strongly augmented labeled source data. Furthermore, a consistency loss is enforced between the student’s predictions on strongly augmented data and the teacher’s predictions on weakly augmented data, applied to samples from both domains. This encourages the student to learn representations invariant to severe perturbations while being guided by more stable targets from the teacher $f_{\mathcal{T}}$. CoMT-VD integrates a dual-strategy contrastive learning (DCL) module that leverages student-extracted features (from strong augmentations) and teacher-extracted features (from weak augmentations) to foster domain-invariant yet discriminative representations. The teacher model updates via exponential moving average (EMA), ensuring training stability and continuous knowledge distillation. This unified framework effectively mitigates cross-domain distribution discrepancies, yielding a violence detection system robust to domain shifts.

3.2.2. Teacher–Student Framework

The student model $f_{\mathcal{S}}(·;{\theta }_{\mathcal{S}})$ learns through direct back-propagation using supervised loss of ${\mathcal{L}}_{sup}$. In contrast, the parameter of the teacher model ${\theta }_{\mathcal{T}}$ are updated as EMA of the student ${\theta }_{\mathcal{T}}$, creating a temporally ensembled, more stable version of the student. The key distinction of CoMT-VD is how data is presented to each model. The student model is trained on aggressively augmented data to learn robust features. In addition, the teacher model, receiving mildly augmented data, provides more consistent and reliable supervisory signals, especially for unlabeled target data and for regularizing the learning process of the student branch.

The student model learns from labeled source data ${\mathit{x}}^{sou}$ via a supervised classification loss ${\mathcal{L}}_{sup}$ (e.g., binary cross-entropy), applied to strongly augmented source samples:

$${\mathcal{L}}_{\sup }={\displaystyle \frac{1}{|{\mathcal{D}}_{sou}|}}\sum _{{\mathit{x}}^{tar}\in {\mathcal{D}}_{sou}}\left[{\mathit{y}}^{sou}logg\left(f_{\mathcal{S}}\left({\mathcal{A}}_{strong}\left({\mathit{x}}^{sou}\right)\right)\right)\right].$$

(1)

Here, ${\overline{\mathit{x}}}_{k,i}^{sou}={\mathcal{A}}_{strong}\left({\mathit{x}}_i^{sou}\right)$ is k-th augmentation from the strong augmenter ${\mathcal{A}}_{strong}$ of i-th sample the source domain. g is the detection head.

Additionally, a consistency loss ${\mathcal{L}}_{cons}$ (e.g., Mean Squared Error) is applied between the predictions of the student on strongly augmented inputs and the teacher on weakly augmented inputs. This loss is computed for all samples ${\mathit{x}}_i$ in a batch, which can include data ${\mathcal{D}}_{tar}$:

$${\mathcal{L}}_{\mathrm{self}}={\displaystyle \frac{1}{|{\mathcal{D}}_{tar}|}}\sum _{x^{tar}\in {\mathcal{D}}_{tar}}\parallel f_{\mathcal{S}}\left({\mathcal{A}}_{strong}\left({\mathit{x}}^{tar}\right)\right)-f_{\mathcal{T}}\left({\mathcal{A}}_{weak}\left({\mathit{x}}^{tar}\right)\right){\parallel }_2^2,$$

(2)

where B is the batch size. This forces the student to produce consistent predictions, even under strong perturbations, aligning with the teacher’s more stable view of the data.

3.2.3. Cross-Domain Augmentation

Integral to the efficacy and robustness of our CoMT-VD framework is a carefully designed cross-domain augmentation pipeline. Far from being a mere data preprocessing step, it serves as a strategic component engineered to (1) fortify the student network ${\theta }_{\mathcal{S}}$ against challenging domain shifts, (2) provide diverse input views essential for dual-strategy contrastive learning to acquire invariant representations, and (3) ensure the stability of supervisory signals generated by the teacher network ${\theta }_{\mathcal{T}}$. To achieve these objectives, we divide the augmentation process into two distinct modalities: a weak augmenter ${\mathcal{A}}_{weak}$, which preserves semantic content to support reliable pseudo-labeling by the teacher, and a strong augmenter ${\mathcal{A}}_{strong}$, which introduces substantial visual variation to enhance the student’s resilience and promote distinctive feature learning.

Weak Augmenter: This is denoted as the operator ${\mathcal{A}}_{weak}$. Its core tenet is the generation of K minimally perturbed view ${\left\{{\tilde{\mathit{x}}}_k\right\}}_{k=1}^K={\mathcal{A}}_{weak}\left(\mathit{x}\right)$, which retains high semantic fidelity to the original sample. This conservative approach is paramount because $\tilde{\mathit{x}}$ serves as the input to the teacher network ${\theta }_{\mathcal{T}}$, whose outputs (i.e., pseudo-labels) should be stable and reliable to guide the student’s optimization on unlabeled data. By minimizing aggressive counterparts, ${\mathcal{A}}_{weak}$ ensures that the teacher’s predictions are not confounded by augmentation-induced artifacts, thereby fostering a more dependable knowledge transfer. We note that ${\mathcal{A}}_{weak}$ concretely includes the following:

• Minor Occlusion: Small, randomly positioned occluding patches are introduced. Given an input $\mathit{x}$ and a binary mask ${\mathit{m}}_{{\tau }_w}$ defining these minor occlusions [22,45], the augmented sample is $\tilde{\mathit{x}}=\mathit{x}\odot (1-{\mathit{m}}_{{\tau }_w})$, where ${\mathit{m}}_{{\tau }_w}\sim \mathrm{Bernoulli}\left({\tau }_w\right)$, and ${\tau }_w$ is set $0.1$ in our experiment.
• Subtle Brightness Adjustment: Pixel intensities are moderately modulated, for example, $\tilde{\mathit{x}}={\gamma }_w·\mathit{x}$, where the brightness factor ${\gamma }_w$ is sampled from a narrow interval of $[0.9,1.1]$.
• Gentle Frame Blending: Subtle temporal alterations or mild blending with non-violent frames ${\mathit{x}}_{nv}$ are performed using a minimal blending coefficient ${\lambda }_w=0.1$, ensuring that the dominant pixel remains unaltered: $\tilde{\mathit{x}}=(1-{\gamma }_w)\mathit{x}+{\gamma }_w{\mathit{x}}_{nv}$.

Given the target domain sample ${\mathit{x}}^{tar}$, the corresponding weak-augmented results ${\left\{{\tilde{\mathit{x}}}_k^{tar}\right\}}_{k=1}^K$ can be obtained:

$${\tilde{\mathit{x}}}_k^{tar}={\mathcal{A}}_{weak}\left({\mathit{x}}^{tar}\right),k=1,2,\dots ,K.$$

(3)

Then, ${\left\{{\tilde{\mathit{x}}}_k^{sou}\right\}}_{k=1}^K$ will be input into the teacher network ${\theta }_T$, to generate high-quality pseudo-labels, which is fundamental for the student’s robust adaptation to the target domain.

Strong Augmenter. In contrast, the strong augmenter ${\mathcal{A}}_{strong}$ is designed to generate substantially diversified and challenging views of the input data. It serves two crucial purposes: (1) It is applied to labeled source samples ${\mathit{x}}^{sou}\in {\mathcal{D}}_{sou}$ to train a student ${\theta }_{\mathcal{S}}$ that is robust to various visual perturbations using self-supervised learning in Equation (2). (2) It is applied to unlabeled target samples ${\mathit{x}}^{tar}\in {\mathcal{D}}_{tar}$ to produce K augmentations. These strong-augmented target data are then processed by the student network ${\theta }_{\mathcal{S}}$ and are the primary inputs for DCL module, compelling it to learn features that are invariant to drastic appearance changes, yet retain semantic discriminability. Concretely, ${\mathcal{A}}_{strong}$ is characterized by significant intensity, broader parameter ranges, as follows:

• Significant Occlusion: Larger or more strategically disruptive occlusion masks ${\mathit{m}}_{{\tau }_s}$ are employed: $\overline{\mathit{x}}=\mathit{x}\odot (1-{\mathit{m}}_{{\tau }_s})$, with  ${\mathit{m}}_{{\tau }_s}\sim \mathrm{Bernoulli}\left({\tau }_s\right)$ and ${\tau }_s=0.3$.
• Major Brightness and Contrast Shifts: Pixel intensities and contrast are altered dramatically. For instance, brightness [46,47] might be scaled by $\overline{\mathit{x}}={\gamma }_s·\mathit{x}$, where ${\gamma }_s$ is sampled from a wider range $[0.5,1.5]$, and contrast adjustments are similarly intensified to simulate challenging real-world lighting conditions (i.e., very dark or overexposed scenes).
• Aggressive Frame Blending and Temporal Manipulation: More profound temporal alterations, such as significant frame shuffling [48], or aggressive blending [49] with disparate scenes ${\mathit{x}}_{nv}$ (including non-violent content or noise) using a substantial blending factor ${\lambda }_s$: $\overline{\mathit{x}}=(1-{\lambda }_s)\mathit{x}+{\lambda }_s{\mathit{x}}_{nv}$, where ${\lambda }_s=0.4$.

The rationale for the strong augmenter is to construct a challenging learning crucible for ${\theta }_{\mathcal{S}}$. We note that the strong augmenter ${\mathcal{A}}_{strong}$ receives the samples ${\mathit{x}}^{tar}$ and ${\mathit{x}}^{sou}$ from both target domain and the source domain, and yields the following conterparts ${\left\{{\overline{\mathit{x}}}^{tar}\right\}}_{k=1}^K$ and ${\left\{{\overline{\mathit{x}}}^{sou}\right\}}_{k=1}^K$:

$$\begin{array}{cc}\hfill {\overline{\mathit{x}}}_k^{tar}& ={\mathcal{A}}_{strong}\left({\mathit{x}}^{tar}\right),k=1,2,\dots ,K,\hfill \end{array}$$

(4)

$$\begin{array}{cc}\hfill {\overline{\mathit{x}}}_k^{sou}& ={\mathcal{A}}_{strong}\left({\mathit{x}}^{sou}\right),k=1,2,\dots ,K.\hfill \end{array}$$

(5)

By forcing the student to discern invariant characteristics across these radically different views of the same underlying instance (especially for target data in the DCL module) and across diverse source instances, we cultivate feature representations that are not only robust to superficial visual changes, but are also highly discriminative of the core semantic content pertaining to violence. This is indispensable for overcoming domain-specific idiosyncrasies and achieving superior generalization.

This strong-weak augmentation strategy [50] establishes a symbiotic interaction within CoMT-VD, effectively addressing domain shift through reliable supervision from the Mean Teacher and DCL’s refinement of discriminative, domain-agnostic features. This synergy forms the foundation of our framework’s enhanced adaptability and performance across diverse violence detection scenarios.

3.2.4. Dual-Strategy Contrastive Learning

To align samples from different distributions, we propose a new contrastive learning mechanism with two key components: Intra-Sample Consistency: Treat a sample and its differently augmented versions as a positive pair, while considering all other samples as negative pairs. Inter-Sample Consistency: To identify positive pairs of the same class from different samples within a batch, we regard samples with similarity scores exceeding ${\delta }^{upper}$ as positive pairs. To better distinguish samples from different classes, samples with similarity scores below ${\delta }^{lower}$ within the same batch are treated as negative pairs. In this Figure, ${\widehat{z}}_k^i$ is the feature map obtained after sample ${\mathit{x}}_k^i$ undergoes the strong augmenter and is fed into the encoder in the student network; ${\overline{\mathit{z}}}_k^i$ is the feature map obtained after sample ${\mathit{x}}_k^i$ undergoes weak augmentation and is fed into the teacher encoder in the teacher network. Here, K represents the number of augmentations. Contrastive learning enables the unsupervised approach of pulling closer the distance between samples of the same category in the input space while repelling samples from different categories [23,24]. While this mechanism is commonly applied in few-shot learning, our work draws inspiration from the core principle of contrastive learning: pulling similar samples closer and pushing dissimilar ones apart in the feature space. As illustrated in Figure 2, we propose a dual positive-negative pair matching strategy to compute the NT-Xent loss (Normalized Temperature-scaled Cross Entropy) [24,51] for both positive and negative sample pairs. We call the module dual-strategy contrastive learning (DCL). The input contains $2K$ augmented samples by the weak augmenter and strong augmenter. These samples first undergo positive-negative pair matching via the Pair Matcher in DCL, followed by the computation of the NT-Xent loss using these pairs to achieve cross-domain sample consistency alignment. The Pair Matcher employs two distinct positive-negative pair matching methods to perform consistency alignment:

• Intra-Sample Consistency: To enable the model to maintain feature invariance of the same sample across different scenarios, we construct positive pairs by pairing the augmented features of a sample with those extracted from other augmentation methods applied to the same sample, while treating features from other samples as negative pairs. This ensures that the model preserves consistency in extracted features across varying environments while enhancing its robustness to noise and perturbations.
• Inter-Sample Consistency: Normal samples consistency establishes positive pairs by matching a sample with its differently augmented views, while treating samples from other instances as negatives—a classic positive-negative pairing strategy. This encourages the network to preserve feature consistency across diverse transformations and real-world variations. However, relying solely on this approach may under-utilize valuable semantic relationships among distinct instances that share the same class label. To fully exploit the diverse information emerging from different views generated by the same sample within a batch, we design an intra-sample feature pairing strategy.

To address this limitation, we introduce Inter-Sample Consistency, which aligns feature representations of different samples belonging to the same class. This not only improves the utilization of intra-class diversity, but also promotes better alignment of heterogeneous yet semantically similar instances, thereby enhancing the compactness and discriminability of action-level features.

As show in Figure 3, we input K strong augmentations ${\left\{{\overline{\mathit{x}}}^{tar}\right\}}_{k=1}^K$ from target domain ${\mathcal{D}}_{tar}$ into the DCL, where each sample ${\overline{\mathit{x}}}_k^{tar}$ is a strong augmentation of target domain using ${\mathcal{A}}_{strong}$. Also, K weak augmentations ${\left\{{\tilde{\mathit{x}}}^{tar}\right\}}_{k=1}^K$ from ${\mathcal{A}}_{weak}$ are input. Then, we encode each ${\overline{\mathit{x}}}_k^{tar}$ by student network $f_{\mathcal{S}}(·;{\theta }_{\mathcal{S}})$, and encode each ${\tilde{x}}_k^{tar}$ by teacher network $f_{\mathcal{T}}(·;{\theta }_{\mathcal{T}})$ to obtain the feature set $\overline{\mathit{Z}}={\left\{\overline{\mathit{z}}\right\}}_{k=1}^K$ and $\tilde{\mathit{Z}}={\left\{\tilde{\mathit{z}}\right\}}_{k=1}^K$. We note that each feature ${\overline{\mathit{z}}}_k$, ${\tilde{\mathit{z}}}_k$ will be used to compute the similarity within the corresponding feature set. Taking a strong augmentation sample pair ${\overline{\mathit{z}}}_m$ and ${\overline{\mathit{z}}}_n$ as an example, the similarity is calculated by cosine similarity:

$$cos\_sim({\overline{\mathit{z}}}_m,{\overline{\mathit{z}}}_n)={\displaystyle \frac{{\overline{\mathit{z}}}_m·{\overline{\mathit{z}}}_n}{\parallel {\overline{\mathit{z}}}_m\parallel \parallel {\overline{\mathit{z}}}_n\parallel }}.$$

(6)

If the $cos\_sim({\overline{\mathit{z}}}_m^i,{\overline{\mathit{z}}}_n^j)>{\delta }^{upper}$, we regard ${\overline{\mathit{z}}}_m$ and ${\overline{\mathit{z}}}_n$ as a positive pair. In contrast, if $cos\_sim({\overline{\mathit{z}}}_m,{\overline{\mathit{z}}}_n)<{\delta }^{lower}$, the ${\overline{\mathit{z}}}_m$ and ${\overline{\mathit{z}}}_n$ are treated as a negative pair. In our experiments, ${\delta }^{upper}=0.7$ is the positive critical threshold and  ${\delta }^{lower}=0.3$ is the negative critical threshold. In the dual-strategy contrastive learning module, for a batch of size K, the teacher network generates K weak augmented samples and the student network generates K strong augmented samples in the unsupervised branch. These 2K samples are then paired as positive and negative sample pairs. All other sample features are disregarded, and we set these similarities to 0. For an augmented feature ${\mathit{z}}_k^{\prime }\in \overline{\mathit{z}}\cup \tilde{\mathit{Z}}$ derived from a sample, its positive pair set is denoted as $\mathcal{P}\left(z_k^{\prime }\right)$. We use the NT-Xent loss [24,51] to compute the contrastive loss ${\mathcal{L}}_{\mathrm{con}}$:

$${\mathcal{L}}_{\mathrm{con}}=-{\displaystyle \frac{1}{2K}}\sum _{k=1}^{2K}log{\displaystyle \frac{{\displaystyle \sum _{{\mathit{z}}_p\in \mathcal{P}\left({\mathit{z}}_k^{\prime }\right)}}exp\left(cos\_sim({\mathit{z}}_k^{\prime },{\mathit{z}}_p)/\tau \right)}{{\displaystyle \sum _{{\mathit{z}}_q\notin \mathcal{P}\left({\mathit{z}}_k^{\prime }\right)}}exp\left(cos\_sim({\mathit{z}}_k^{\prime },{\mathit{z}}_q)/\tau \right)}},$$

(7)

where $\tau$ is the temperature scaling factor. We are consistent with the best performance in [24], and set $\tau$ = 0.1. $\mathcal{P}\left(z_k^{\prime }\right)$ represent positive pair sets for the ${\mathit{z}}_k^{\prime }$, and ${\mathit{z}}_p$ represents the positive feature for ${\mathit{z}}_k^{\prime }$ in the $\mathcal{P}\left(z_k^{\prime }\right)$, and ${\mathit{z}}_q$ represents the negative sample of $\mathcal{P}\left(z_k^{\prime }\right)$, which is not in P and have a similarity ${\delta }^{lower}<0.3$ with ${\mathit{z}}_k^{\prime }$.

3.2.5. Mean Teacher Optimization

The Mean Teacher architecture enhances semi-supervised learning by leveraging the stability of the teacher model ${\theta }_{\mathcal{T}}$ to reduce noise in pseudo-labels [12]. To ensure that the teacher network acquires richer knowledge while maintaining stability, the weakly augmented samples ${\tilde{\mathit{x}}}_k^{tar}={\mathcal{A}}_{weak\left(x^{tar}\right)}$ are fed into the teacher branch. This approach enables the model to learn sample features across diverse scenarios while preserving the teacher model’s stability. As a result, the teacher generates more reliable pseudo-labels to guide the parameters updating of the student network ${\theta }_{\mathcal{S}}$. The parameters ${\theta }_{\mathcal{S}}$ of student network $f_{\mathcal{S}}(·;{\theta }_{\mathcal{S}})$ is updated using gradient descent:

$${\theta }_{\mathcal{S}}\leftarrow {\theta }_{\mathcal{S}}-\alpha ·{\nabla }_{{\theta }_{\mathcal{S}}}{\mathcal{L}}_{\mathrm{CoMT}},$$

(8)

where $\alpha =0.001$ is the learning rate. In addition, ${\mathcal{L}}_{\mathrm{CoMT}}$ is the total loss function of the CoMT-VD framework, which is formulated as the combination of three components, i.e., the supervised loss ${\mathcal{L}}_{\sup }$, self-supervised loss ${\mathcal{L}}_{\mathrm{self}}$, and contrastive loss ${\mathcal{L}}_{\mathrm{con}}$ in Equation (7), expressed as

$${\mathcal{L}}_{\mathrm{CoMT}}={\lambda }_1{\mathcal{L}}_{\sup }+{\lambda }_2{\mathcal{L}}_{\mathrm{self}}+{\lambda }_3{\mathcal{L}}_{\mathrm{con}},$$

(9)

where ${\lambda }_1=1$, ${\lambda }_2=0.5$, and ${\lambda }_3=0.3$ are the weights of these three loss terms.

Meanwhile, the parameters ${\theta }_{\mathcal{T}}$ of the teacher branch are updated via ${\theta }_{\mathcal{S}}$ by Exponential Moving Average (EMA):

$${\theta }_{\mathcal{T}}\leftarrow \eta {\theta }_{\mathcal{T}}+(1-\eta ){\theta }_{\mathcal{S}}.$$

(10)

Here, $\eta$ is the momentum factor, which is recommended to set to $0.95$ during deployment. The overall training procedure of the proposed CoMT-VD (Contrastive Mean Teacher for Violence Detection) model is outlined in Algorithm 1.

Algorithm 1 Training Procedure of CoMT-VD

Require: Labeled source data ${\mathcal{D}}_{sou}=\left\{({\mathit{x}}_i^{sou},y_i^{sou})\right\}$; Unlabeled target data ${\mathcal{D}}_{tar}=\left\{{\mathit{x}}^{tar}\right\}$; Pre-trained model $f_{\theta }$; Output: Optimized teacher $f_{{\theta }_{\mathcal{T}}}$ and student model $f_{{\theta }_{\mathcal{S}}}$; 1. Initialize the student network $f_{{\theta }_{\mathcal{S}}}\leftarrow f_{\theta }$ and the teacher network $f_{{\theta }_{\mathcal{T}}}\leftarrow f_{\theta }$; 2. For each ${\mathit{x}}_i^{sou}\in {\mathcal{D}}_{sou}$ and ${\mathit{x}}_i^{tar}\in {\mathcal{D}}_{tar}$ do: ⊳ generate the strong augmentations 3.     ${\left\{{\overline{\mathit{x}}}_k^{sou}\right\}}_{k=1}^K={\mathcal{A}}_{strong}\left(x^{sou}\right)$; 4.     ${\left\{{\overline{\mathit{x}}}_k^{tar}\right\}}_{k=1}^K={\mathcal{A}}_{strong}\left(x^{tar}\right)$; ▹ generate the weak augmentations 5.     ${\left\{{\tilde{\mathit{x}}}_k^{tar}\right\}}_{k=1}^K={\mathcal{A}}_{weak}\left(x^{tar}\right)$; ⊳ compute the feature of the strong augmented samples 6.    ${\left\{\overline{\mathit{z}}\right\}}_k^K=f_{{\theta }_{\mathcal{T}}}\left({\left\{{\overline{\mathit{x}}}_k^{tar}\right\}}_{k=1}^K\right)$ and ${\left\{\tilde{\mathit{z}}\right\}}_k^K=f_{{\theta }_{\mathcal{T}}}\left({\left\{{\tilde{x}}_k^{tar}\right\}}_{k=1}^K\right)$; 7.    Calculate ${\mathcal{L}}_{\sup }$ using Equation (1);      ${\mathcal{L}}_{\sup }={\displaystyle \frac{1}{|{\mathcal{D}}_{sou}|}}{\sum }_{{\mathit{x}}^{tar}\in {\mathcal{D}}_{sou}}\left[y^{sou}logg\left(f_{\mathcal{S}}\left({\mathcal{A}}_{strong}\left({\mathit{x}}^{sou}\right)\right)\right)\right]$; 8.     Calculate ${\mathcal{L}}_{\mathrm{self}}$ via Equation (2); 9.     Calculate ${\mathcal{L}}_{\mathrm{con}}$ via Equation (7); 10.   Calculate ${\mathcal{L}}_{\mathrm{CoMT}}={\mathcal{L}}_{\sup }+{\mathcal{L}}_{\mathrm{self}}+{\mathcal{L}}_{\mathrm{con}}$; 11.   Update ${\theta }_{\mathcal{S}}$ using Equation (8); 12.   Update ${\theta }_{\mathcal{T}}$ using Equation (10); 13. End For

4. Experiments

We implement our proposed model based on TensorFlow 2.3.0, and all experiments are deployed on a machine with a 24 GB NVIDIA A10 GPU. In this section, we firstly introduce our experimental datasets, then we will present the experimental setup and results of the model. In this paper, all of the experiments adopted the classic frame sampling method in the violence detection [3,6,7,8], sampling 16 frames uniformly from each video. In this paper, all training runs were conducted with the random seed fixed at 42, and the best-performing model was selected after 200 epochs of training. Following this, we conduct a further ablation analysis, and, finally, we analyze our proposed model based on the experimental results.

4.1. Benchmark Datasets

In this paper, we verify the performance of our proposed model on five well-known datasets:

• RLVS [7] is a large-scale real-world captured dataset, containing 1000 violent and 1000 non-violent video clips. The video data in RLVS features more diverse scenarios including streets, classrooms, courtyards, corridors, and sports fields. Most footage is captured from a third-person perspective, predominantly showing fights involving multiple individuals, with a small number of group violence scenes.
• RWF-2000 [52] is designed for real-world violent behavior detection under surveillance cameras. The videos in this dataset are collected from raw surveillance footage on YouTube, segmented into clips of up to 5 s at 30 frames per second, with each clip labeled as either violent or non-violent behavior.
• Hockey Fight [53] comprises 1000 real video clips from ice hockey matches, evenly split between 500 fight scenes and 500 normal segments in the hockey games. However, a notable limitation of this dataset is the lack of scene diversity, as all videos are confined to ice hockey rinks during matches.
• Movies [53] is relatively small-scale, consisting of 200 video clips from action movies, with 100 clips depicting violent scenes and 100 non-violent scenes.
• Violent Flow [54] describes violent or non-violent group behaviors in real-world scenarios. The samples originates from surveillance cameras and monitoring devices, capturing large-scale crowd behaviors in public spaces such as stadiums, streets, and squares. The videos are divided into two categories, violent video and normal video, with a total of 246 video clips (123 violent and 123 non-violent clips).

4.2. Baselines

Our CoMT-VD framework can be integrated with any existing violence detection models. Therefore, to validate the generalization capability of our proposed CoMT-VD, we conduct experiments on 5 different models categorized into three groups based on their architectures: (1) 3DCNN-based: Song et al. [55] extends the C3D network proposed in [12] (originally designed for human action recognition) to the task of violent behavior recognition. Zhenhua et al. [56] proposes full-temporal fusion violence detection model (SCTF). (2) 2DCNN-based: Soliman et al. [7] uses 2DCNN to extract spatial features from each frame and employs an LSTM network for temporal feature fusion, called VGG16 + LSTM. Wang et al. [6] proposes a plug-and-play module (ActionNet) for human action recognition networks, which is also used to the violence detection task. (3) Transformer-based: Arnab et al. [57] introduces a pure Transformer encoder for video understanding, replacing conventional 3D convolutions in action recognition tasks.

4.3. Experimental Setup

To comprehensively evaluate the efficacy of our proposed CoMT-VD framework, we designed two distinct experimental setups, each addressing a critical aspect of real-world violence detection challenges: handling unseen environmental conditions and generalizing to variational datasets. In all experiments, inputs consist of 16 video frames uniformly segmented from the video, with each frame resized to a resolution of $224\times 224$.

(1) Setup-1: Performance under Simulated Variational Conditions: This setup is meticulously designed to assess CoMT-VD’s robustness and generalization abilities when confronted with data corrupted by various challenging, yet common, environmental factors not explicitly seen during standard training. Our objective is to simulate real-world unseen scenarios to validate the performance improvement of CoMT-VD on data exhibiting variational distributions. We introduce four challenging conditions, occlusion, fog, rain, and low-light environments, as shown in Figure 4.

• Rain Simulation: Gaussian noise is strategically injected into the original images, with elongated noise points designed to simulate realistic raindrops. Raindrop parameters are precisely set to a length of 10 pixels and a count of 500 per frame, ensuring a consistent and reproducible simulation of rainy scenes [16].
• Fog Simulation: A fog-like color template [16] (fixed at an RGB value of 200) is overlaid onto the original image at 50% intensity. This process emulates the reduced visibility and contrast of foggy conditions.
• Low-Light Simulation: The brightness of each frame is systematically reduced to 40–70% of its original level. This simulates varying degrees of low-light conditions, a frequent challenge in surveillance and monitoring applications.
• Occlusion Simulation: To mimic real-world occlusions [22,45], randomly positioned rectangular masks, with widths and heights sampled from the range of $[30,70]$ pixels, are applied to each frame within the test videos. This simulates partial obstructions that can severely impact visual cues.

For this setup, five baseline models are evaluated across five benchmark datasets. For each dataset, 80% of the data is allocated for training and 20% for testing. Crucially, within our proposed CoMT-VD, the training set is further partitioned, with 20% designated as source domain data and 80% as target domain data. We conduct a rigorous comparative analysis of baseline model both before and after the integration of CoMT-VD to empirically validate its effectiveness in mitigating the impact of these unseen conditions.

(2) Setup-2: Generalization to Variational Datasets: This experimental setup focuses on evaluating the cross-scenario generalization efficacy. Here, the five baselines are initially trained on widely recognized real-world violence detection datasets, namely RLVS [7] and RWF-2000 [52]. Subsequently, their performance is rigorously evaluated on distinct, scenario-specific benchmarks: Hockey Fight [53] and Movies [53]. To ensure comparative validity and a fair assessment, baseline counterparts augmented with our proposed training strategy undergo identical evaluation protocols. This testing paradigm quantitatively assesses the detection accuracy improvements, thereby systematically validating the methodology’s ability to generalize adequately to entirely variational datasets, a critical requirement for real-world applicability.

In all experiments presented in this paper, the input consists of 16 video frames uniformly segmented from the video, with each frame resized to a resolution of $224\times 224$.

4.4. Experimental Results

4.4.1. Performance for Unseen Conditions

In real-world scenarios (such as the video obtained from security cameras without modification by multimedia technology), there are often uncertain dynamic environmental changes (e.g., occlusion, low illumination, rainy, foggy weather, etc.) that involve unseen data not encountered during training, leading to performance degradation of models. This experiment aims to validate the capability of our proposed CoMT in handling such unseen datasets under environmental variations. We train five baseline models on five benchmark datasets, respectively, using 80% of the data for training and 20% for testing. The test data simulates challenging scenarios including rainy, foggy, low-light, and occluded conditions to evaluate the models’ generalization capability on unseen situations potentially encountered in real-world deployments. To validate the effectiveness of the Contrastive Mean Teacher (CoMT), we integrate CoMT into each of the five pre-trained baseline models for fine-tuning, subsequently testing their adaptability and robustness in previously unseen extreme scenarios.

Table 1. Volence detection performance (accuracy %) across five dataset under different domain shift scenarios: rainy, foggy, low-light, and occlusion scenes. The numbers in parentheses (blue color) indicate the absolute performance gains (percentage points) achieved by CoMT-VD over vanilla baseline models.

Methods	VGG16 + LSTM [7]	VGG16 + LSTM + CoMT-VD	C3D [55]	C3D + CoMT-VD	ViViT [57]	ViViT + CoMT-VD	ActionNet [6]	ActionNet + CoMT-VD	SCTF [56]	SCTF + CoMT-VD
ACC
Datesets
	Rainny Scenes
RLVS [7]	71.5	78.3 (+6.8)	70.3	78.5 (+8.2)	69.5	79.3 (+9.8)	73.3	80.3 (+7.0)	74.0	82.8 (+8.8)
RWF-2000 [52]	69.5	78.8 (+9.3)	70.8	79.3 (+8.5)	68.8	76.3 (+7.5)	70.0	79.5 (+9.5)	73.5	79.8 (+6.3)
Hockey Fight [53]	72.3	83.5 (+9.2)	71.5	82.3 (+10.8)	70.5	78.0 (+7.5)	73.5	82.3 (+8.8)	76.3	85.0 (+9.7)
Movies [53]	75.0	80.0 (+5.0)	72.5	80.0 (+7.5)	72.5	77.5 (+5.0)	75.0	80.0 (+5.0)	77.5	85.0 (+7.5)
Violent Flow [54]	72.0	84.0 (+12.0)	76.0	82.0 (+6.0)	70.0	78.0 (+8.0)	76.0	86.0 (+10.0)	78.0	84.0 (+6.0)
	Foggy Scenes
RLVS [7]	71.8	77.5 (+5.7)	69.3	76.3 (+7.0)	69.8	75.0 (+5.2)	74.5	84.5 (+10.0)	73.0	83.3 (+10.3)
RWF-2000 [52]	67.3	78.5 (+11.2)	69.8	78.3 (+8.5)	69.5	75.8 (+6.3)	72.3	80.3 (+8.0)	73.8	81.5 (+7.8)
Hockey Fight [53]	69.5	78.0 (+8.5)	71.0	77.5 (+6.5)	68.0	77.0 (+9.0)	73.0	82.0 (+9.0)	75.5	83.5 (+8.0)
Movies [53]	72.5	77.5 (+5.0)	69.0	80.0 (+11.0)	68.5	77.5 (+9.0)	75.0	85.0 (+10.0)	75.0	82.5 (+7.5)
Violent Flow [54]	70.0	76.0 (+6.0)	72.0	80.0 (+8.0)	68.0	78.0 (+10.0)	78.0	84.0 (+6.0)	78.0	86.0 (+8.0)
	Low-light Scenes
RLVS [7]	70.0	75.5 (+5.5)	69.3	77.0 (+7.7)	67.5	76.8 (+9.3)	70.3	81.3 (+11.0)	72.0	81.8 (+9.8)
RWF-2000 [52]	67.0	79.3 (+12.3)	68.5	78.8 (+9.3)	66.5	75.3 (+8.8)	69.0	79.0 (+10.0)	71.3	79.3 (+8.0)
Hockey Fight [53]	69.0	78.5 (+9.5)	68.0	78.0 (+10.0)	67.5	77.0 (+9.5)	71.5	80.5 (+9.0)	75.5	82.0 (+6.5)
Movies [53]	70.0	77.5 (+7.5)	70.0	77.5 (+7.5)	70.0	75.0 (+5.0)	70.0	80.0 (+10.0)	75.0	82.0 (+7.0)
Violent Flow [54]	68.0	78.0 (+10.0)	72.0	80.0 (+10.0)	70.0	76.0 (+6.0)	76.0	86.0 (+10.0)	76.0	84.0 (+8.0)
	Occlusion Scenes
RLVS [7]	68.8	76.8 (+8.0)	70.3	78.8 (+8.5)	67.5	77.3 (+9.8)	71.3	82.5 (+11.2)	72.0	84.3 (+12.3)
RWF-2000 [52]	67.3	78.5 (+11.2)	68.3	79.0 (+10.7)	66.8	77.0 (+11.2)	72.0	80.3 (+10.3)	71.8	81.3 (+9.5)
Hockey Fight [53]	68.5	79.0 (+10.5)	69.0	80.5 (+11.5)	67.5	77.5 (+10.0)	72.5	81.0 (+8.5)	75.5	82.5 (+7.0)
Movies [53]	70.0	77.5 (+7.5)	67.5	80.0 (+12.5)	67.5	77.5 (+10.0)	72.5	80.0 (+7.5)	72.5	82.5 (+10.0)
Violent Flow [54]	66.0	78.0 (+12.0)	68.0	76.0 (+8.0)	66.0	76.0 (+10.0)	78.0	84.0 (+6.0)	76.0	86.0 (+10.0)

presents the performance comparisons between the baseline models and those integrated with CoMT-VD across five benchmark datasets for rainy scenarios, foggy scenarios, low-light scenarios, and occlusion scenarios, respectively. It can be observed that the incorporation of CoMT-VD leads to a notable improvement in the model’s performance on unseen scenarios. We conduct the experiment in the settings consistent with Setup-1. The results show that the introduction of CoMT-VD can enhance the generalization capability of the models, improving its detection performance on unseen data. The baseline models achieve improvements ranging from 5.0∼12.0%, 6.0∼12.5%, 5.0∼11.2%, 5.0∼11.2%, and 6.3∼12.3% across the five benchmark datasets in four different domain scenarios, respectively. We note that the results demonstrate that CoMT-VD consistently enhances the capability of various violence detection models for test samples with variational distributions.

4.4.2. Performance for Cross Datasets

To validate CoMT’s ability to enhance model generalization and improve domain adaptation capability, we designed this experiment to demonstrate that models trained on real-world datasets (RLVS, RWF-2000) show improved domain adaptation and generalization on specific-scenario violence detection datasets (Hockey Fight, Movies). The experimental setup during training remained consistent with Section 4.4.1, using real-world datasets RLVS and RWF individually, as well as a mixed dataset combining RLVS and RWF-2000 for training. Testing was conducted on the Hockey Fight and Movies datasets. The Hockey Fight dataset contains violent behaviors occurring in ice hockey arenas, where all subjects wear protective gear. While the dataset exhibits notable variations in scene configurations and subject appearances, the characteristic patterns of violent behaviors remain consistent. The Movies dataset contains numerous exaggerated violent actions from a movie perspective, which differs from the distribution of daily violent behaviors in the training data. As shown in

Table 2. Performance (accuracy %) of models trained on real-world datasets test on the Hockey Fight dataset under Setup-1. We employ real-world datasets (RLVS and RWF-2000, both of them come from surveillance videos, which include indoor/outdoor fights and daily violent behaviors) as training data and evaluated on the Hockey Fight dataset to validate the model’s generalization capability for cross-domain detection. The numbers in parentheses (blue color) indicate the absolute performance gains (percentage points) achieved by CoMT-VD over vanilla baseline models.

Methods	VGG16 + LSTM	VGG16 + LSTM + CoMT-VD	C3D	C3D + CoMT-VD	ViViT	ViViT + CoMT-VD	ActionNet	ActionNet + CoMT-VD	SCTF	SCTF + CoMT-VD
ACC
Datesets
RLVS	67.4	75.2 (+7.8)	69.3	75.1 (+5.8)	68.7	74.8 (+6.1)	72.9	78.6 (+5.7)	74.1	79.5 (+5.4)
RWF-2000	68.3	75.8 (+7.5)	70.0	76.7 (+6.7)	69.4	78.6 (+9.2)	74.9	80.0 (+5.1)	73.5	81.7 (+8.2)
RLVS + RWF-2000	71.3	78.7 (+7.4)	72.3	79.5 (+7.2)	71.3	78.9 (+7.6)	75.8	82.9 (+7.1)	75.0	82.3 (+7.3)

and

Table 3. Performance of models trained on real-world datasets test on the Movies dataset under Setup-2. We employ real-world datasets (RLVS and RWF-2000, both of them come from surveillance videos, which include indoor/outdoor fights and daily violent behaviors) as training data and evaluated on the Movies dataset to validate the model’s generalization capability for cross-domain detection. The numbers in parentheses (blue color) indicate the absolute performance gains (percentage points) achieved by CoMT-VD over vanilla baseline models.

Methods	VGG16 + LSTM	VGG16 + LSTM + CoMT-VD	C3D	C3D + CoMT-VD	ViViT	ViViT + CoMT-VD	ActionNet	ActionNet + CoMT-VD	SCTF	SCTF + CoMT-VD
ACC
Datesets
RLVS	72.5	78.0 (+5.5)	71.5	76.5 (+5.0)	69.5	76.0 (+6.5)	76.0	82.0 (+6.0)	74.5	82.0 (+7.5)
RWF-2000	69.0	77.5 (+8.5)	72.0	79.5 (+7.5)	68.0	74.0 (+6.0)	74.0	83.5 (+9.5)	76.0	84.0 (+8.0)
RLVS + RWF-2000	72.0	80.0 (+8.00)	70.0	82.0 (+12.0)	72.0	78.0 (+6.0)	78.0	86.5 (+8.5)	76.5	86.0 (+9.5)

, baseline models incorporating the CoMT strategy achieve remarkable performance improvements. The results indicate that these baselines can enhance cross-domain detection capability when trained with CoMT-VD, specifically improving the models trained on real-world data by 5.1∼9.2% on the Hockey Fight test set and 5.0∼12.0% on the test set of the Movies dataset. Experimental results demonstrate that incorporating CoMT adequately enhances the model’s generalization performance in cross-domain detection under such specialized scenarios

5. Ablation Studies

Finally, we investigate the impact of different components on the generalization capability of the CoMT-VD model. When we set the brightness ${\gamma }_s=[0.5,1.5]$ ${\gamma }_w=[0.9-1.1]$, occlusion ${\tau }_w=0.1$ ${\tau }_s=0.3$, Frame Blending ${\lambda }_s=0.4$ ${\lambda }_w=0.1$, and similarity threshold ${\delta }^{upper}=0.7$ ${\delta }^{lower}=0.3$ parameters, we fine-tuned the model to achieve the best performance. In this section, we introduce the impact of different parameter settings and modules on the model.

5.1. Different Augmentation Intensities

This study employs three distinct data augmentation strategies to imbue training data with multi-scenario noise perturbations, thereby enhancing the model’s adaptability to diverse environmental disturbances. The experiments investigate how varying augmentation intensities affect the model’s detection generalization capability: (1) Brightness Augmentation: We optimize brightness augmentation factors ${\gamma }_s$ and ${\gamma }_w$ through systematic strength calibration to identify optimal parameter values. (2) Occlusion Augmentation: By controlling occlusion ratios ${\mathit{m}}_{{\tau }_s}$ and ${\mathit{m}}_{{\tau }_w}$, selecting diverse occlusion box dimensions, we enhance the model’s noise resistance against partial occlusions. The numbers of pixels with occlusion are calculated by $\lfloor {\mathit{m}}_{{\tau }_s}\odot \mathit{x}\rfloor \odot \lfloor {\mathit{m}}_{{\tau }_s}\odot \mathit{x}\rfloor$, $\lfloor {\mathit{m}}_{{\tau }_w}\odot \mathit{x}\rfloor \odot \lfloor {\mathit{m}}_{{\tau }_w}\odot \mathit{x}\rfloor$, respectively, where $\lfloor ·\rfloor$ represents rounding down to an integer. Ther weak augmenter randomly samples occlusion masks as integers within (0, [$\lfloor {\mathit{m}}_{{\tau }_w}\odot \mathit{x}\rfloor$], while the strong augmentation selects integers from [$\lfloor {\mathit{m}}_{{\tau }_w}\odot \mathit{x}\rfloor$,${\mathit{m}}_{{\tau }_s}\odot \mathit{x}\rfloor$]. (3) Temporal Frame Blending: A frame blending rate parameter governs the insertion of adversarial frames into video sequences. Given input dimensions $224\times 224\times 16$, the number of blended frames is calculated as $\lfloor {\lambda }_{nv}\times 16\rfloor$, strengthening temporal feature extraction capabilities. In this section, all experiments maintain consistency with Setup-2 in terms of scenario configurations and data usage. During the testing phase, we amalgamate the four scenarios defined in Setup-2 into composite test data for experimental validation.

Impact of Brightness Adjustment Factor. As demonstrated in

Table 4. Performance (accuracy %) with different brightness augmentations. The numbers of red color represent the absolute performance degradation (percentage) lossed by CoMT-VD compared to the best performance. The first column specifies different value combinations of illumination ${\gamma }_s$ (strong augmentation) and ${\gamma }_w$ (weak augmentation) in the illumination-reduction augmentation.

Methods		RLVS [7]	RWF-2000 [52]	Hockey Fight [53]	Movies [53]	Violent Flow [54]
ACC
${\mathit{\gamma }}_{\mathit{s}}$ ${\mathit{\gamma }}_{\mathit{w}}$
[0.5, 1.5]	[0.9–1.1]	81.8	79.3	82.3	81.5	84.0
	[0.9–1.2]	79.0 (−2.8)	80.1 (−2.2)	79.6 (−1.9)	78.8 (−2.7)	80.2 (−3.8)
	[1.0–1.1]	78.4 (−3.4)	79.2 (−3.1)	78.1 (−4.2)	77.8 (−3.7)	78.8 (−5.2)
	[0.8–1.0]	79.5 (−2.3)	76.3 (−3.0)	78.5 (−3.8)	77.0 (−4.5)	78.00 (−6.0)
[0.5, 1.4]	[0.9–1.1]	77.2 (−4.6)	76.4 (−2.9)	78.0 (−4.3)	76.5 (−5.0)	79.3 (−4.7)
[0.6, 1.5]		76.8 (−5.0)	75.0 (−4.3)	79.6 (−1.9)	77.8 (−3.7)	78.2 (−5.8)
[0.4, 1.5]		78.0 (−3.8)	78.8 (−2.7)	79.3 (−2.2)	78.2 (−3.3)	80.1 (−3.9)

, we explored varying random sampling ranges for the brightness augmentation factor. Optimal results are achieved by configuring strong augmentation with a factor randomly sampled from $[0.5,1.5]$ and weak augmentation from $[0.9,1.1]$. This reveals two key observations: A clear distinction exists between the ranges of strong and weak brightne augmentation factors. Weak augmentation operates within a narrower, more stable range, whereas strong augmentation notably enhances the model’s generalization capability when employing a broader yet appropriately bounded random sampling interval. The empirical observations demonstrate that both excessive or insufficient degradation levels adversely impact model generalization capability. The bold font indicates the best performance achieved by CoMT-VD across all setups.

Impact of Occlusion Augmentation Rate. As shown in

Table 5. Performance (accuracy %) with different occlussion augmentation rates in the strong and weak augmenter. The numbers of red color represent the absolute performance degradation (percentage) lossed by CoMT-VD compared to the best performance. This experiment investigates the impact of occlusion size on model generalization capability. The numbers of pixels with occlusion are calculated by $\lfloor {\mathit{m}}_{{\tau }_s}\odot \mathit{x}\rfloor \odot \lfloor {\mathit{m}}_{{\tau }_s}\odot \mathit{x}\rfloor$,$\begin{array}{c}\hfill break\end{array}$$\lfloor {\mathit{m}}_{{\tau }_w}\odot \mathit{x}\rfloor \odot \lfloor {\mathit{m}}_{{\tau }_w}\odot \mathit{x}\rfloor$ respectively, where $\lfloor ·\rfloor$ represents rounding down to an integer.

Methods		RLVS [7]	RWF-2000 [52]	Hockey Fight [53]	Movies [53]	Violent Flow [54]
ACC
${\mathit{\tau }}_{\mathit{w}}$ ${\mathit{\tau }}_{\mathit{s}}$
0.3	0.1	81.8	79.3	82.3	81.5	84.0
	0.2	77.8 (−4.0)	76.3 (−3.0)	79.8 (−2.5)	78.5 (−3.0)	79.7 (−4.3)
0.2	0.1	76.2 (−5.6)	75.0 (−4.3)	77.5 (−4.8)	76.3 (−5.2)	79.1 (−5.9)
0.4		78.3 (−3.5)	77.0 (−5.3)	78.6 (−3.7)	78.3 (−3.2)	80.0 (−4.0)

, the model achieves optimal performance when the occlusion ratio is set to 0.3 for strong augmentation and 0.1 for weak augmentation. Excessively high occlusion ratios (i.e., ${\tau }_s>0.3$ in strong augmentation) may excessively mask critical features during training, impairing feature extraction capabilities. Conversely, overly conservative ratios (e.g., lowering strong augmentation below ${\tau }_s<0.3$) risk oversimplifying learned representations, thereby degrading generalization ability. Crucially, insufficient differentiation between strong and weak augmentation parameters (e.g., $0.2$ for weak vs. $0.3$ for strong and $0.1$ for weak vs. $0.2$) leads to dual failures: inadequate knowledge acquisition due to overlapping perturbation intensities and diminished pseudo-label reliability caused by weakened data credibility in weakly augmented samples.

Impact of Temporal Frame Blending Rate: This experiment investigates the impact of introducing a cross-modal data mixing strategy within the Mean Teacher framework. In

Table 6. Performance (%) with different frame blending rate. The numbers of red color represent the absolute performance degradation (percentage) lossed by CoMT-VD compared to the best performance. The number of blended frames is calculated as $\lfloor {\lambda }_{nv}\times 16\rfloor$, strengthening temporal feature extraction capabilities.

Methods		RLVS [7]	RWF-2000 [52]	Hockey Fight [53]	Movies [53]	Violent Flow [54]
ACC
${\mathit{\lambda }}_{\mathit{s}}$ ${\mathit{\lambda }}_{\mathit{w}}$
0.4	0.1	81.8	79.3	82.3	83.5	84.0
	0.2	79.3 (−2.5)	76.3 (−3.0)	78.6 (−3.7)	80.6 (−2.9)	80.2 (−3.8)
0.3	0.1	78.5 (−3.3)	78.2 (−1.1)	77.9 (−4.4)	80.6 (−2.7)	79.3 (−4.7)
0.5		77.3 (−4.5)	74.5 (−3.8)	77.2 (−5.1)	79.2 (−4.3)	79.8 (−4.2)

, by comparing the effects of inserting non-violent frames at different ratios in the strong/weak augmentation branches (${\lambda }_s=0.3$ for strong augmentation and ${\lambda }_w=0.1$ for weak augmentation), we observe that introducing 30% non-violent frames in the strong augmentation branch achieves optimal validation performance. This demonstrates the strategy’s effectiveness in enhancing the model’s discriminative capability against adversarial frames. The high-disturbance features generated by strong augmentation, combined with the semantic interference from non-violent frames, create an adversarial training paradigm that compels the network to prioritize temporal motion features over individual frame appearances. Meanwhile, maintaining a lower 10% mixing ratio in the weak augmentation branch preserves the reliability of pseudo-labels generated by the teacher network. It can be observed that excessive blending frames rate may lead to the replacement of key feature frames, which could degrade model performance. Insufficient distinction between strong and weak augmentations might result in inadequate knowledge acquisition by the model, while unstable weak augmentations could compromise pseudo-label quality and further impair model performance. In this study, we select the parameter combination of and ${\lambda }_s=0.4$ and ${\lambda }_w=0.1$.

5.2. The Impact of Other Factors on Performance

Impact of Mean Teacher and DCL in CoMT-VD: In this experiment, the encoder employs two baselines: the top-performing SCTF and the classic 3DCNN-based model C3D. We validate the generalization of CoMT-VD on unseen data across five benchmark datasets, specifically the models are tested on the Hockey Fight dataset, to demonstrate CoMT-VD’s capability. Following the experimental setups in Section 4.4.1, we evaluate the impact of CoMT-VD’s components by disabling the Mean Teacher strategy and DCL strategy separately across different baselines.

Table 7. Impact of accuracy by Mean Teacher and DCL for cross-dataset detection. To verify the contributions of both modules, we conduct experiments by selectively disabling each component through setting their corresponding loss terms to zero, thereby validating the effectiveness of both components in CoMT-VD. These experiments are performed by training on three datasets (RLVS, RWF-2000, and RLVS + RWF-2000) and testing on the Hockey Fight dataset, aiming to demonstrate CoMT-VD’s capability in addressing domain shift challenges for cross-dataset detection tasks. The numbers in parentheses (blue color) indicate the absolute performance gains (percentage points) achieved by CoMT-VD over vanilla baseline models. × indicates that the corresponding module has been included, while √ indicates that the corresponding module has not been included.

Methods			Dataset
Baseline	Mean Teacher	DCL	RLVS [7]	RWF-2000 [52]	RLVS + RWF-2000
C3D	×	×	69.3	70.0	72.3
	√	×	73.3 (+4.0)	75.5 (+5.50)	76.0 (+3.7)
	×	√	71.4 (+2.1)	74.5 (+2.1)	74.25 (+1.75)
	√	√	75.1 (+5.8)	76.7 (+6.7)	79.5 (+7.2)
SCTF	×	×	74.1	73.5	75.0
	√	×	76.3 (+2.2)	78.0 (+4.5)	78.3 (+3.3)
	×	√	76.2 (+2.1)	76.5 (+6.0)	77.5 (+2.5)
	√	√	79.5 (+5.4)	81.7 (+8.2)	82.3 (+7.3)

,

Table 8. Impact of F1 score by Mean Teacher and DCL for cross-dataset detection. The numbers in parentheses (blue color) indicate the absolute performance gains (percentage points) achieved by CoMT-VD over vanilla baseline models. × indicates that the corresponding module has been included, while √ indicates that the corresponding module has not been included.

Methods			Dataset
Baseline	Mean Teacher	DCL	RLVS [7]	RWF-2000 [52]	RLVS + RWF-2000
C3D	×	×	0.709	0.715	0.713
	√	×	0.728 (+0.019)	0.749 (+0.034)	0.742 (+0.029)
	×	√	0.716 (+0.007)	0.734 (+0.021)	0.729 (+0.013)
	√	√	0.748 (+0.039)	0.756 (+0.041)	0.757 (+0.044)
SCTF	×	×	0.712	0.724	0.728
	√	×	0.731 (+0.019)	0.732 (+0.008)	0.738 (+0.010)
	×	√	0.728 (+0.016)	0.729 (+0.005)	0.736 (+0.008)
	√	√	0.769 (+0.057)	0.773 (+0.049)	0.786 (+0.058)

,

Table 9. Validation of the Impact of accuracy by Mean Teacher and DCL on the Models for unseen scenarios dataset violent video detection. We conduct experiments by selectively disabling each component through setting their corresponding loss terms to zero, thereby validating the effectiveness of both components in CoMT-VD. These experiments are performed by training on five benchmark datasets (RLVS, RWF-2000, Hockey Fight, Movies and Violent) respectively, and test the performance on each simulated scenarios of test set to demonstrate CoMT-VD’s capability in addressing domain shift challenges for unseen conditions dataset detection tasks. The numbers in parentheses (blue color) indicate the absolute performance gains (percentage points) achieved by CoMT-VD over vanilla baseline models. × indicates that the corresponding module has been included, while √ indicates that the corresponding module has not been included.

Methods			Dataset
Baseline	Mean Teacher	DCL	RLVS [7]	RWF-2000 [52]	Hockey Fight [53]	Movies [53]	Violent Flow [54]
C3D	×	×	70.3	69.4	70.6	71.6	72.0
	√	×	74.0 (+3.7)	75.3 (+5.9)	76.0 (+5.4)	78.5 (+7.5)	76.0 (+3.3)
	×	√	72.5 (+2.2)	73.8 (+4.4)	74.4 (+3.8)	72.5 (+4.5)	70.0 (+4.0)
	√	√	77.3 (+7.0)	78.8 (+10.3)	79.5 (+11.0)	79.0 (+11.0)	78.0 (+12.0)
SCTF	×	×	72.6	72.8	75.5	75.7	76.2
	√	×	79.5 (+6.9)	78.0 (+5.2)	80.0 (+4.5)	79.5 (+3.8)	82.60 (+6.4)
	×	√	75.5 (+2.9)	74.6 (+1.8)	78.3 (+2.8)	77.3 (+1.6)	79.3 (+3.1)
	√	√	83.8 (+11.2)	80.2 (+7.4)	82.7 (+6.3)	81.5 (+5.8)	85.7 (+9.5)

and

Table 10. Validation of the Impact of F1 score by Mean Teacher and DCL on the Models for unseen scenarios dataset violent video detection. The numbers in parentheses (blue color) indicate the absolute performance gains (percentage points) achieved by CoMT-VD over vanilla baseline models. × indicates that the corresponding module has been included, while √ indicates that the corresponding module has not been included.

Methods			Dataset
Baseline	Mean Teacher	DCL	RLVS [7]	RWF-2000 [52]	Hockey Fight [53]	Movies [53]	Violent Flow [54]
C3D	×	×	0.712	0.721	0.723	0.719	0.726
	√	×	0.726 (+0.014)	0.737 (+0.016)	0.742 (+0.019)	0.751 (+0.032)	0.748 (+0.022)
	×	√	0.719 (+0.007)	0.729 (+0.008)	0.738 (+0.015)	0.727 (+0.008)	0.740 (+0.014)
	√	√	0.732 (+0.019)	0.743 (+0.022)	0.748 (+0.025)	0.756 (+0.037)	0.751 (+0.025)
SCTF	×	×	0.723	0.730	0.739	0.732	0.741
	√	×	0.743 (+0.020)	0.755 (+0.025)	0.762 (+0.023)	0.786 (+0.054)	0.791 (+0.050)
	×	√	0.736 (+0.013)	0.742 (+0.012)	0.769 (+0.030)	0.751 (+0.019)	0.773 (+0.032)
	√	√	0.802 (+0.079)	0.781 (+0.051)	0.793 (+0.054)	0.810 (+0.078)	0.797 (+0.056)

demonstrate that incorporating the Mean Teacher and dual-strategy contrastive learning (DCL) modules can enhance generalization by improving the effectiveness of training samples. While DCL alone—without the Mean Teacher framework—yields moderate gains in generalization, the combination of both components achieves the best performance, highlighting their complementary roles in facilitating cross-domain knowledge transfer. In this experiment, we aggregate unseen data from diverse categories across each dataset into the test set to evaluate the model’s out-of-distribution performance. Figure 5 illustrates the ROC curves of C3D and SCTF in cross-dataset violence detection after training on three source datasets with CoMT-VD. The results show that CoMT-VD improves the Area Under the ROC Curve (AUC), indicating enhanced model performance in balancing true positive and false positive rates when detecting violent behavior.

Impact of the threshold in DCL: Within the Inter-sample consistency in the proposed DCL, our pair-matching mechanism for positive/negative samples considers feature pairs with similarities above threshold ${\delta }^{upper}$ as positive matches (features from different samples but belonging to the same class), and those below threshold ${\delta }^{lower}$ as negative matches (features from different samples with distinct classes). This approach ensures that pseudo-labels generated in unlabeled self-supervised learning better approximate the true class distribution. The selection of these pair-matching thresholds becomes crucial. The results in

Table 11. Performance (accuracy %) with different similarity thresholds in DCL. The numbers of red color represent the absolute performance degradation (percentage) lossed by CoMT-VD compared to the best performance. For the proposed DCL, the inter-sample consistency alignment controls the matching of positive and negative pairs through upper-bound ${\delta }^{upper}$ and lower-bound ${\delta }^{lower}$. From the result, we observe the combination of $0.7$ and $0.3$ is the optimal values.

Methods		RLVS [7]	RWF-2000 [52]	Hockey Fight [53]	Movies [53]	Violent Flow [54]
ACC
${\mathit{\delta }}^{\mathit{upper}}$ ${\mathit{\delta }}^{\mathit{lower}}$
0.7	0.3	81.8	79.3	82.3	81.5	84.0
	0.4	79.0 (−2.8)	78.5 (−2.8)	79.6 (−2.7)	79.3 (−2.2)	82.6 (−1.4)
	0.2	78.9 (−2.9)	79.0 (−0.3)	79.8 (−2.5)	78.9 (−2.6)	80.9 (−3.1)
0.6	0.3	79.8 (−2.0)	78.2 (−1.1)	79.7 (−1.6)	79.2 (−2.3)	82.2 (−1.8)
0.8		78.3 (−3.5)	77.3 (−2.6)	80.2 (−2.1)	80.4 (−1.1)	81.2 (−2.8)

indicate that the optimal performance is achieved when ${\delta }^{upper}$ = 0.7 and ${\delta }^{lower}$ = 0.3. Excessively strict constraints on inter-sample similarity matching can lead to insufficient alignment capability learned from the data, hindering the model’s knowledge acquisition. Conversely, overly lenient constraints may introduce excessive invalid sample pairs, causing the model to learn irrelevant or even counterproductive knowledge.

Impact of the weights in the loss function. Equation (9) represents the loss function of CoMT-VD. Since labeled data typically provides more reliable optimization for models in supervised settings, we set ${\lambda }_1=1.0$. This experiment aims to explore the effects of varying ${\lambda }_2$ and ${\lambda }_3$ on model performance and identify the optimal combination of ${\lambda }_2$ and ${\lambda }_3$ for CoMT-VD. We integrate CoMT-VD into the pretrained SCTF model and trained it on RLVS, RWF-2000, and their combined dataset (RLVS + RWF-2000). By fine-tuning the model with different combinations and evaluating accuracy on the Hockey Fight and Movies datasets, we select the best parameter configuration via observing the accuracy trend with different ${\lambda }_2$ and ${\lambda }_3$. Figure 6 demonstrate that the model achieves optimal performance when ${\lambda }_2=0.5$ and ${\lambda }_3=0.3$.

Performance of other augmentation: In this ablation study, we evaluate classical image augmentation techniques, rotation and cropping, finding they slightly reduce model performance, as summarized in

Table 12. Validation of the performance (accuracy %) of other augmentations. The numbers in parentheses (red color) represent the absolute performance degradation (percentage) loss by CoMT-VD compared to the best performance sets.

Augmentation Methods			Dataset
Augmenter in CoMT-VD	Rotarion	Cropping	RLVS [7]	RWF-2000 [52]	Hockey Fight [53]	Movies [53]	Violent Flow [54]
Occlusion Brightness Blending	×	×	83.8	80.2	82.7	81.5	85.7
	×	√	81.6 (−2.2)	79.3 (−0.9)	81.5 (−1.2)	79.0 (−2.5)	83.6 (−2.1)
	√	×	80.3 (−3.5)	79.1 (−1.1)	80.9 (−1.8)	81.8 (−1.9)	84.5 (−1.2)

. For rotation, we implement ${30}^{\circ }$ (strong) and ${15}^{\circ }$ (weak) angles. For cropping, randomly selected dimensions are used with ratios $\beta =0.6$ (strong) and $\beta =0.8$ (weak). The performance decrease can be attributed to the inherent translation invariance of CNNs, where rotation fails to alter feature relationships adequately. Furthermore, unlike occlusion, cropping inadvertently obscures critical continuous features in sequential data, leading to inferior performance. This suggests that inappropriate augmentation causes models to learn ineffective features, indicating an upper bound to the knowledge gain from augmented data.

6. Conclusions

6.1. Discussion

In this paper, we address the critical challenge of domain shift in automated violence detection systems, which notably impedes their reliable deployment in diverse real-world surveillance scenarios. We introduce CoMT-VD, a new Contrastive Mean Teacher-based Violence Detection model, specifically engineered to enhance model adaptability and robustness across unseen target domains. The efficacy of CoMT-VD stems from its innovative integration of a Mean Teacher architecture, which adequately leverages unlabeled target domain data to foster stable, domain-invariant feature representations through consistency regularization. Crucially, to counteract the potential supervisory noise from pseudo-labels and further refine the learned feature space, we incorporate a sophisticated dual-strategy contrastive learning module. It meticulously refines features by enforcing both intra-sample consistency for robust representation learning and inter-sample consistency for sharpened categorical discriminability. Broad evaluations on five benchmark datasets unequivocally demonstrate the superior capacity to substantially improve generalization performance on unseen data and across challenging cross-dataset assessments, marking a notable advancement towards more robust and reliable real-world violence detection systems.

6.2. Limitation

This paper verifies the proposed training strategy for enhancing the model’s ability to handle domain shift problems. Due to the limitation of difficulty in collecting violent behavior datasets in specific scenarios, this study employs the addition of various types of noise to general open source datasets for violent behavior detection, simulating common real-world conditions such as low-light, rainy, foggy, and occluded scenes (as illustrated in Figure 4). However, whether the proposed training strategy can maintain consistent performance under corresponding real-world environmental changes requires further validation in subsequent research using datasets from actual specific scenarios. Additionally, this paper focuses on fine-tuning pre-trained models to improve model performance, but it has limitations in detecting continuous video sequences with fewer than 16 frames, necessitating special processing techniques (such as padding video frames or implementing looped playback). Furthermore, according to the ablation experiments, the model has limitations in detecting videos where the number of frames containing violent behavior is extremely low.

6.3. Future Work

For future work, we plan to explore several promising directions, including the integration of more advanced spatiotemporal backbone networks to further enhance feature extraction capabilities, investigating adaptive weighting mechanisms for dual consistency losses or other self-supervised learning forms tailored for adversarial domain shifts, extending CoMT-VD to a multi-modal framework incorporating audio cues for richer contextual understanding, and optimizing CoMT-VD for real-time inference and deployment on edge devices to address computational efficiency.