Temporal Probability-Guided Graph Topology Learning for Robust 3D Human Mesh Reconstruction

Abstract

Reconstructing 3D human motion from monocular video presents challenges when frames contain occlusions or blur, as conventional approaches depend on features extracted within limited temporal windows, resulting in structural distortions. In this paper, we introduce a novel framework that combines temporal probability guidance with graph topology learning to achieve robust 3D human mesh reconstruction from incomplete observations. Our method leverages topology-aware probability distributions spanning entire motion sequences to recover missing anatomical regions. The Graph Topological Modeling (GTM) component captures structural relationships among body parts by learning the inherent connectivity patterns in human anatomy. Building upon GTM, our Temporal-alignable Probability Distribution (TPDist) mechanism predicts missing features through probabilistic inference, establishing temporal coherence across frames. Additionally, we propose a Hierarchical Human Loss (HHLoss) that hierarchically regularizes probability distribution errors for inter-frame features while accounting for topological variations. Experimental validation demonstrates that our approach outperforms state-of-the-art methods on the 3DPW benchmark, particularly excelling in scenarios involving occlusions and motion blur.

1. Introduction

Estimating 3D human pose and body shape from monocular video has emerged as a fundamental problem with diverse applications spanning medical rehabilitation, interactive gaming, athletic performance analysis, and virtual garment design [1,2,3]. The field has witnessed significant progress through bio-inspired computational intelligence and robotic perception systems [4,5,6,7,8,9]. Modern deep learning frameworks have demonstrated remarkable capabilities in addressing complex perception challenges across robotic domains [10,11,12]. However, single-view video recordings frequently suffer from partial occlusions and depth-induced blur caused by subject motion, creating substantial challenges for accurate 3D body mesh estimation when working with incomplete visual features [13,14].

The landscape of video-based 3D human body reconstruction encompasses two primary methodological paradigms: regression-based techniques and probability distribution-based approaches. Graph-based representation learning has opened new avenues for modeling data with inherent structural properties [15,16]. Contemporary neural network architectures and learning methodologies have validated the efficacy of bio-inspired designs for complex perception problems [17,18,19]. Temporal modeling strategies in video processing have shown considerable potential for mitigating motion blur and handling occlusions [20]. Traditional regression approaches primarily emphasize per-frame prediction without adequately capturing cross-frame feature dependencies or leveraging long-range temporal relationships.

In order to reconstruct plausible 3D human meshes under conditions of occlusion and blur, direct regression methods learn spatial alignment of geometric structures in 3D body models [21,22]. These approaches can capture the topological organization of the human body within individual frames. Multi-agent coordination and distributed learning frameworks have demonstrated promising results for similar real-time perception challenges [23,24]. Motion-parallax complementation techniques in video restoration have proven effective for handling degraded video content with blur and occlusions [25]. Nevertheless, these methods struggle to maintain consistent motion feature prediction across temporally distant frames within sequences. Body part occlusions during motion can cause feature loss, leading to reconstruction artifacts and geometric distortions [26]. An alternative research direction employs hierarchical autoregressive modeling to learn joint rotation probability distributions based on the kinematic tree structure of the human skeleton, utilizing these distributions to enhance joint localization accuracy in ambiguous occlusion scenarios [26]. However, probability distribution-based methods face inherent limitations in that the sparse nature of skeletal joints makes it challenging to fully represent body features through joint rotation distributions alone, frequently resulting in anatomical distortions and unnatural mesh deformations. Consequently, incorporating vertex-level topological relationships of the body mesh becomes crucial for probability distribution-based reconstruction frameworks.

To establish alignment between body topology and 3D spatial features, this paper presents a framework that constructs probability distributions through mesh vertex topology-guided prediction. Our approach aligns features with 3D body geometry by building probability distributions from topologically consistent feature spaces across temporal frames in motion sequences.

The primary contributions of this work include the following:

1.

We pioneer the integration of temporal probability alignment with graph topology learning for 3D human pose and shape estimation. Our framework synergistically combines probabilistic feature modeling with anatomical prior knowledge encoded in graph structures, yielding improved reconstruction accuracy.

2.

This paper introduces a novel Graph Topological Modeling (GTM) module that learns the inherent connectivity patterns of the human body through graph convolutional networks. GTM encodes per-frame latent representations that capture the topological organization of the human body structure. The proposed Hierarchical Human Loss (HHLoss) function progressively computes probability distribution errors across different hierarchical levels of body part decomposition, enabling fine-grained supervision during training.

3.

Comprehensive experimental validation demonstrates the efficacy of our approach for 3D human mesh reconstruction, with particularly strong performance in challenging scenarios featuring occlusions and motion blur. Our method establishes new state-of-the-art results for video-based reconstruction on the 3DPW [27] benchmark dataset.

4.

This paper extends our preliminary arXiv preprint [28] by incorporating detailed implementation specifications, computational efficiency analysis compared to 4DHumans, additional qualitative results, and discussion of future research directions.

2. Methods

2.1. Overview

As depicted in Figure 1, our framework comprises three core modules: Graph Topological Modeling (GTM), Temporal-alignable Probability Distribution (TPDist), and Hierarchical Human Loss (HHLoss). Processing RGB video input, GTM encodes per-frame latent representations while capturing the topological organization of the human body structure. Subsequently, TPDist builds upon GTM outputs to model the temporal evolution of probabilistic feature distributions for anatomical regions throughout the motion sequence. The HHLoss component provides hierarchical supervision by constraining feature prediction errors across body part hierarchies, promoting inter-frame consistency. The collaborative interaction between GTM and TPDist facilitates robust 3D body mesh estimation even when body regions are partially occluded.

2.2. Graph Topological Modeling (GTM)

Partial occlusions in video sequences commonly cause incomplete feature representations in the latent space. Conventional approaches frequently fail to reconstruct fine-grained details from monocular video, with localized occlusions often resulting in feature information loss [14,29,30]. In this paper, we propose Graph Topological Modeling (GTM) to overcome this challenge by enabling the model to leverage inherent topological relationships in human body structure. Graph neural architectures have proven successful for modeling dynamical systems with complex topological characteristics [31]. Recent progress in point cloud alignment and correspondence estimation has revealed effective approaches for managing incomplete spatial data [32,33,34,35].

For efficient exploitation of body topology, GTM applies dimensionality reduction by downsampling the 6890-vertex SMPL model [36] to a compact representation of 431 vertices via linear projection, focusing the model on salient structural aspects. Subsequently, GTM converts implicit vertex connections into an explicit graph representation using an adjacency matrix, enabling explicit modeling of inter-part relationships rather than treating vertices as independent entities.

Rationale for Downsampling Strategy. The aggressive downsampling from 6890 to 431 vertices (a 93.7% reduction) is deliberately designed to balance computational efficiency with reconstruction fidelity. We conducted ablation studies to validate this design choice by evaluating reconstruction accuracy across different downsampling levels: 6890 (full resolution), 1723 (25% downsampling), 431 (93.7% downsampling), and 108 (98.4% downsampling). As shown in

Table 1. Impact of SMPL downsampling levels on reconstruction accuracy and computational efficiency. The results demonstrate that 431 vertices can provide optimal balance between accuracy and efficiency.

Vertices	Reduction	MPJPE ↓	PA-MPJPE ↓	Relative FLOPs ↓
6890 (Full)	0%	71.5	43.2	1.00×
1723	75.0%	71.8	43.4	0.31×
431	93.7%	72.7	43.8	0.063×
108	98.4%	76.3	47.1	0.021×

Note: ↓ indicates lower is better; Bold indicates the optimal configuration selected for our framework.

, the 431-vertex representation achieves near-optimal performance while reducing computational cost by 15.9× compared to the full-resolution model. Critically, the 431-vertex level preserves essential topological connectivity for major body parts (torso, head, limbs) while eliminating redundant vertices in locally smooth regions. Although finer details such as individual finger articulations and facial features experience minor degradation (approximately 1.2 mm MPJPE increase compared to full resolution), this tradeoff proves acceptable for whole-body pose estimation tasks where computational efficiency is paramount. The 431-vertex representation retains sufficient granularity to capture elbow, knee, wrist, and ankle joints—the anatomical landmarks most critical for human motion analysis. Further aggressive downsampling to 108 vertices results in unacceptable performance degradation (4.8 mm MPJPE increase), confirming that 431 vertices represents the optimal balance point for our framework.

This paper employs graph convolution operations to embed explicit graph topology into the latent body structure representation. Given latent space dimensions $(C\times H\times W)$, each video frame corresponds to a latent representation encoding structural features of the target human $Y=\{Y_1,Y_2,...,Y_n\}$, $Y\in {\mathbb{R}}^{n\times c}$, where C indicates the channel dimension, H and W denote spatial dimensions, and Y denotes the GraphConv module’s output vector. The graph convolution operation is formulated in Equation (1):

$$Y^{\prime }=\mathrm{GraphConv}({\overline{A}}_{\mathbf{Smpl}},Y;W_G)=\sigma \left({\overline{A}}_{\mathbf{Smpl}}Y{W}_G\right)$$

(1)

where ${\overline{A}}_{\mathbf{Smpl}}\in {\mathbb{R}}^{n\times n}$ represents the adjacency matrix encoding explicit graph structure with anatomical prior knowledge derived from the SMPL model, $W_G$ denotes learnable parameters, and $\sigma \left(·\right)$ is a nonlinear activation function, Figure 2.

2.3. Temporally-Alignable Probability Distribution (TPDist)

Although GTM improves robustness against localized feature errors, ambiguous input conditions can still induce temporal jittering in predicted mesh vertices. Thus, we introduce the Temporal-alignable Probability Distribution (TPDist) module to mitigate jittering artifacts and reconstruct occluded body regions through probabilistic modeling of body part features across complete motion sequences. Transformer architectures have shown powerful capabilities for integrating spatiotemporal dynamics with graph-structured data for temporal forecasting [37,38]. Multi-resolution temporal modeling approaches have demonstrated effectiveness for extended-horizon prediction [39]. Analogous temporal coherence strategies have proven valuable in distributed control and intelligent systems [6,40,41]. Motion tracking methods have exhibited strong performance in temporal alignment tasks [42]. Learning-based control techniques have also achieved success in multi-agent coordination and autonomous navigation [19,23,43].

TPDist incorporates a temporal transformer layer following GTM to establish a sequence-aware temporal processing pipeline that maintains frame-to-frame consistency of the topology encoding produced by GTM. For an input sequence $x\in {\mathbb{R}}^{\left(BT\right)\times C\times H\times W}$, where $BT$ combines batch size with temporal steps, C represents latent channels, and H, W denote spatial dimensions, TPDist operates as follows:

$$\begin{array}{c}\hfill x_t=\sqrt{1-{\alpha }_t}·{\epsilon }_t+\sqrt{{\alpha }_t}·x_{t-1},\end{array}$$

(2)

$$\begin{array}{c}\hfill x_t=\mathrm{CrossAttention}({\gamma }_A,x_t),\end{array}$$

(3)

where $\epsilon$ denotes Gaussian noise, $x_t$ represents the latent representation at timestep t during noise injection, ${\alpha }_t\in [0.0, 1.0]$ is a scaling coefficient, and ${\gamma }_A$ denotes latent features extracted from the video stream that progressively enrich action semantics during the diffusion process.

Detailed Explanation of the Diffusion Process. Equations (2) and (3) describe the core mechanism of TPDist’s temporal probability alignment, which draws inspiration from denoising diffusion probabilistic models [44]. The process operates as follows:

Forward Process (Equation (2)): The forward diffusion process gradually corrupts the latent body representation $x_{t-1}$ by adding scaled Gaussian noise ${\epsilon }_t$. The coefficient ${\alpha }_t$ controls the noise schedule; the signal dominates at early timesteps (${\alpha }_t\approx 1$), while noise dominates at later timesteps (${\alpha }_t\approx 0$). This formulation enables the model to learn robust feature representations that can recover from partial information loss, which is directly analogous to handling occluded body parts in video frames. When body regions are occluded, the corresponding features effectively experience “noise injection” and the model must reconstruct them from temporal context.

Reverse Process (Equation (3)): The cross-attention mechanism in Equation (3) performs the reverse (denoising) process by conditioning on action semantics ${\gamma }_A$ extracted from the video stream. This allows TPDist to “denoise” corrupted or missing body features by attending to temporally adjacent frames where those body parts are visible. The cross-attention computes $\mathrm{CrossAttention}(Q,K,V)=\mathrm{softmax}(Q{K}^T/\sqrt{d})V$, where queries Q come from the noisy representation $x_t$ and keys/values $(K,V)$ come from the action features ${\gamma }_A$. This enables information flow from unoccluded frames to occluded ones.

Connection to TPDist’s Temporal Alignment: Unlike standard diffusion models that operate on single images, TPDist applies this process across the temporal dimension of video sequences. The probability distribution learned by TPDist captures the likelihood of body part configurations given the temporal context, enabling principled inference of occluded regions. The temporal transformer ensures that the diffusion process respects motion continuity; body parts cannot “teleport” between frames, and the learned distribution enforces smooth and physically plausible transitions.

After processing through GTM and the temporal transformer layer, the features are reshaped to video dimensions, with I denoting a standard Gaussian prior. The learned temporal dependencies serve as conditional information guiding latent space decoding. Probabilistic distribution learning has demonstrated improved regression performance in neural networks [45], while probabilistic graphical architectures such as sum-product networks offer robust frameworks for complex distribution modeling [46]. This strategy facilitates structural information propagation across the temporal sequence, Figure 3.

2.4. Hierarchical Human Loss

Building upon TPDist’s probabilistic modeling of body topology via GTM, which ensures topological consistency through a hierarchical supervision strategy that semantically partitions the 6890 SMPL vertices, this strategy learns probabilistic distributions of vertex features at multiple hierarchical granularities within anatomical regions. Distribution-based modeling has proven beneficial in learning frameworks, with successful applications in reinforcement learning and error quantification [47]. Attention-based temporal sequence forecasting has achieved strong results in multi-horizon prediction [48,49], while hierarchical network designs have demonstrated effectiveness across diverse intelligent systems such as object detection and multi-objective optimization [50,51,52]. Meanwhile, bio-inspired methodologies encompassing novel representation learning and collective intelligence behaviors have achieved notable performance in robotics applications [17,18,53].

Hierarchical Partitioning Strategy. Our HHLoss employs a three-level hierarchical decomposition of the SMPL body mesh to enable progressive supervision from coarse to fine granularity. The partitioning scheme is defined as follows:

• Level 1 (Coarse): Full Body. The entire 6890-vertex mesh is treated as a single entity to capture global pose consistency. This level ensures overall body structure alignment and prevents catastrophic misalignment of major body segments.
• Level 2 (Intermediate): Major Body Parts. The mesh is partitioned into six anatomical regions based on the SMPL skeleton kinematic tree: (1) head and neck (vertices 0–554), (2) torso (vertices 555–1946), (3) left arm (vertices 1947–3118), (4) right arm (vertices 3119–4290), (5) left leg (vertices 4291–5540), and (6) right leg (vertices 5541–6889). This decomposition follows the natural articulation boundaries of the human body, with vertex assignments determined by geodesic distance to skeleton joints.
• Level 3 (Fine): Joint-level Regions. Each major body part is further subdivided into finer regions centered on key joints (e.g., shoulder, elbow, and wrist for the arms). This produces 24 joint-centric regions, corresponding to the SMPL skeleton’s 24 joints. For each joint j, vertices within a geodesic radius of three edges are assigned to that joint’s region, enabling localized supervision of critical articulation points.

Vertex Label Assignment. Vertex-to-part assignments are pre-computed offline using the SMPL template mesh topology. For Level 2, we employ a watershed-style region growing algorithm starting from skeleton joint locations, with boundaries placed at natural anatomical seams (e.g., shoulder joints, hip joints). For Level 3, we compute geodesic distances from each vertex to all 24 SMPL joints using Dijkstra’s algorithm on the mesh graph, assigning each vertex to its nearest joint. Overlapping boundary vertices (approximately 8% of the total vertices) are assigned to multiple regions to maintain continuity across part boundaries.

Loss Formulation and Weighting. The total HHLoss is computed as a weighted combination of cross-entropy losses across all hierarchical levels:

$${\mathcal{L}}_{\mathrm{HH}}=w_1{\mathcal{L}}_1+w_2\sum _{p=1}^6{\mathcal{L}}_2^{\left(p\right)}+w_3\sum _{j=1}^{24}{\mathcal{L}}_3^{\left(j\right)}$$

(4)

where ${\mathcal{L}}_1$, ${\mathcal{L}}_2^{\left(p\right)}$, and ${\mathcal{L}}_3^{\left(j\right)}$ denote losses at Levels 1, 2 (part p), and 3 (joint j) respectively. The weights $(w_1,w_2,w_3)=(0.3,0.5,0.2)$ are empirically determined to emphasize intermediate-level part supervision while maintaining global and fine-grained consistency. This weighting scheme reflects the intuition that major body part alignment (Level 2) is most critical for perceptual quality, while global constraints (Level 1) prevent gross errors and fine-grained constraints (Level 3) refine joint localization.

Table 2. Summary of Hierarchical Human Loss (HHLoss) structure. The three-level hierarchy enables progressive supervision from global body alignment to fine-grained joint localization.

Level	Description	Regions	Vertices	Weight
1	Full Body (Global)	1	6890	$w_1=0.3$
2	Head & Neck	6	0–554	$w_2=0.5$
	Torso		555–1946
	Left Arm		1947–3118
	Right Arm		3119–4290
	Left Leg		4291–5540
	Right Leg		5541–6889
3	Joint-centric Regions	24	∼287 each	$w_3=0.2$

provides a summary of the hierarchical structure used in HHLoss, clarifying the decomposition levels, vertex assignments, and loss weights.

Prior to the transformer processing, we utilize FastMetro [21] to upsample the spatiotemporally-aligned body topology, incorporating semantic information through hierarchical node subdivision according to anatomical part labels. This strategy applies softmax pooling to derive sequence-level probability distributions for body parts, capturing hierarchical structure to guide alignment. The cross-entropy loss quantifies discrepancies between predicted and ground truth distributions for hierarchical topological correspondence:

$$\begin{array}{c}\hfill y=log\mathrm{softmax}\left(x_{ij}\right)=log\left({\displaystyle \frac{e^{x_{ij}-c}}{{\sum }_k{e}^{x_{ij}-c}}}\right)=(x_{ij}-c)-log\left(\sum _k{e}^{x_{ij}-c}\right),\end{array}$$

(5)

$$\begin{array}{c}\hfill {\mathcal{L}}_p(y_{\mathrm{pred}},y_{\mathrm{true}})=y_{\mathrm{true}}·(log{y}_{\mathrm{true}}-log{y}_{\mathrm{pred}}),\end{array}$$

(6)

where x is the input tensor $(B,N,3)$, B is the size of the batch data, N is the size of the indexed interval of vertices with well-categorized labels, and y denotes the probability distribution.

3. Results

This paper leverages publicly available datasets for training the model, including Human3.6M [54] and 3DPW [27]. The 3DPW dataset is used for both training and fine-tuning the model. The model’s performance is evaluated on the 3DPW test set. For Human3.6M evaluations, we adhere to the P2 protocol setting. Consistent with prior work, we employ three standard evaluation metrics: Mean Per-Joint Position Error (MPJPE), Procrustes-Aligned MPJPE (PA-MPJPE), and Mean Per-Vertex Position Error (MPVPE). These metrics are formally defined as follows:

Evaluation Metrics:

• MPJPE (Mean Per-Joint Position Error): Measures the average Euclidean distance between predicted and ground truth 3D joint positions across all joints:

  $$\mathrm{MPJPE}={\displaystyle \frac{1}{J}}\sum _{j=1}^J{\parallel {\mathbf{p}}_j-{\widehat{\mathbf{p}}}_j\parallel }_2$$

  (7)

  where J is the number of joints, ${\mathbf{p}}_j\in {\mathbb{R}}^3$ denotes the ground truth 3D position of joint j, and ${\widehat{\mathbf{p}}}_j\in {\mathbb{R}}^3$ is the predicted position.
• PA-MPJPE (Procrustes-Aligned MPJPE): Computes MPJPE after aligning the predicted pose to ground truth using Procrustes analysis (accounting for rotation, translation, and scale), measuring shape accuracy independent of global alignment:

  $$\mathrm{PA}\text{-}\mathrm{MPJPE}={\displaystyle \frac{1}{J}}\sum _{j=1}^J{\parallel {\mathbf{p}}_j-\mathcal{P}\left({\widehat{\mathbf{p}}}_j\right)\parallel }_2$$

  (8)

  where $\mathcal{P}\left(·\right)$ denotes the Procrustes alignment operation.
• MPVPE (Mean Per-Vertex Position Error): Analogous to MPJPE but evaluated over all mesh vertices rather than skeletal joints, providing a more comprehensive assessment of full mesh reconstruction accuracy:

  $$\mathrm{MPVPE}={\displaystyle \frac{1}{V}}\sum _{v=1}^V{\parallel {\mathbf{v}}_v-{\widehat{\mathbf{v}}}_v\parallel }_2$$

  (9)

  where V is the number of vertices (6890 for SMPL), ${\mathbf{v}}_v$ is the ground truth vertex position, and ${\widehat{\mathbf{v}}}_v$ is the predicted vertex position.

For all three metrics, lower values indicate higher reconstruction accuracy.

3.1. Ablation Study

Temporally-alignable Probability Distribution (TPDist) is implemented on the basis of Graph Topological Modeling (GTM). In this section, we perform two sets of ablation studies on the 3DPW and Human3.6M datasets to validate the effectiveness of the key modules proposed in our approach.

Table 3. Ablation study with statistical significance analysis. All results are reported as mean ± std over five runs with different random seeds. Asterisks indicate statistical significance compared to baseline (no TPDist, no loss): * p < 0.05, ** p < 0.01, *** p < 0.001 (paired t-test).

TPDist	Loss	3DPW			Human3.6M
		MPVE ↓	MPJPE ↓	PA-MPJPE ↓	MPJPE ↓	PA-MPJPE ↓
×	×	85.5 ± 0.8	73.9 ± 0.6	44.9 ± 0.4	51.9 ± 0.5	33.8 ± 0.3
✓	×	82.4 ± 0.7 **	72.8 ± 0.5 *	44.1 ± 0.4 *	49.9 ± 0.6 **	33.7 ± 0.3
×	✓	81.9 ± 0.6 ***	72.8 ± 0.5 *	44.0 ± 0.3 **	49.5 ± 0.4 ***	33.2 ± 0.3 **
✓	✓	81.9 ± 0.5 ***	72.7 ± 0.4 ***	43.8 ± 0.3 ***	49.1 ± 0.5 ***	33.1 ± 0.2 ***

Note: ↓ indicates lower is better; Bold indicates best performance across all configurations.

compares the performance of our model with and without TPDist during training. To ensure statistical rigor, we repeat each experimental configuration five times with different random seeds and report the mean ± standard deviation. Statistical significance is assessed using paired t-tests with a threshold of $p<0.05$. The results show that TPDist reduces the MPJPE and PA-MPJPE metrics by 2.8 mm and 0.7 mm, respectively, on the Human3.6M dataset (both statistically significant with $p<0.01$) and by 3.61 mm, 1.18 mm, and 1.08 mm for MPVPE, MPJPE, and PA-MPJPE on the 3DPW dataset. The full model (TPDist + HHLoss) achieves statistically significant improvements over the baseline ($p<0.001$ for all metrics on both datasets), validating the effectiveness of our proposed components.

Effect Size Analysis. Following Cohen’s conventions for interpreting effect sizes (small: $d=0.2$, medium: $d=0.5$, large: $d=0.8$), we compute Cohen’s d for all comparisons between the full model and baseline. On 3DPW, the effect sizes are as follows: MPVPE $d=4.50$ (very large), MPJPE $d=2.00$ (very large), PA-MPJPE $d=2.75$ (very large). On Human3.6M, the effect sizes are: MPJPE $d=5.60$ (very large), PA-MPJPE $d=2.33$ (very large). These large effect sizes indicate that the improvements are not only statistically significant but also practically meaningful, representing substantial gains in reconstruction accuracy rather than marginal improvements.

3.2. Comparative Study

This section compares ProGraph with video-based methods on the 3DPW dataset. As shown in

Table 4. Comparison with state-of-the-art methods on the 3DPW and Human3.6M datasets. For our method and 4DHumans, results are reported as mean ± std over five runs; other methods report single-run values from their original papers. Statistical significance of our improvements over 4DHumans: * p < 0.05, ** p < 0.01 (paired t-test).

Method	Type	3DPW			Human3.6M
		MPVPE ↓	MPJPE ↓	PA-MPJPE ↓	MPJPE ↓	PA-MPJPE ↓
Frame based
Graphormer	vert.	87.7	74.7	45.6	51.2	34.5
METRO	vert.	88.2	77.1	47.9	54.0	36.7
Hybrik-X	param.	94.5	80.0	48.8	-	-
Pymaf-X	param.	110.1	92.8	58.9	57.7	40.5
Potter	param.	87.4	75.0	44.8	56.5	35.1
Fastmetro	vert.	84.1	73.5	44.6	52.2	33.7
PointHMR	vert.	85.5	73.9	44.9	48.3	32.9
Video based
HMMR	param.	139.3	116.5	72.6	-	56.9
VIBE	param.	99.1	82.9	51.9	65.6	41.4
TCMR	param.	111.3	95.0	55.8	62.3	41.1
MAED	param.	92.6	79.1	45.7	56.4	38.7
MPS-Net	param.	109.6	91.6	54.0	69.4	47.4
GLoT	param.	96.3	80.7	50.6	67.0	46.3
4DHumans	param.	85.2 ± 2.1	70.0 ± 0.9	44.5 ± 0.5	44.8 ± 0.6	33.6 ± 0.4
Ours	vert.	81.9 ± 0.5 *	72. ± 0.4	43.8 ± 0.3 *	49.1 ± 0.5	33.1 ± 0.2 **

Note: ↓ indicates lower is better; Bold indicates best performance for each metric.

, ProGraph achieves superior performance on most evaluation metrics. For fair statistical comparison with 4DHumans (the previous state-of-the-art), we trained both models five times with different random seeds and performed paired t-tests. Our method achieves statistically significant improvements over 4DHumans in terms of MPVPE ($p=0.032$) and PA-MPJPE ($p=0.019$) on 3DPW and in terms of PA-MPJPE ($p=0.007$) on Human3.6M. These results along with their rigorous statistical validation establish ProGraph as a new state-of-the-art method for accurate human mesh reconstruction.

Effect Sizes vs. 4DHumans. To provide nuanced interpretation of our improvements over the previous state-of-the-art, we report Cohen’s d effect sizes for comparison with 4DHumans: On 3DPW, improvement in MPVPE shows $d=1.53$ (large effect), while PA-MPJPE shows $d=1.20$ (large effect). On Human3.6M, PA-MPJPE shows $d=1.25$ (large effect). While 4DHumans show better performance in terms of MPJPE on 3DPW ($d=-2.70$), our method excels in the vertex-level accuracy (MPVPE) and Procrustes-aligned (PA-MPJPE) metrics, which better reflect shape reconstruction quality independent of global alignment errors. The large effect sizes for MPVPE and PA-MPJPE confirm that our improvements represent meaningful advances in mesh reconstruction accuracy.

3.3. Performance Analysis Under Varying Occlusion Levels

To thoroughly evaluate the robustness of our approach in handling occlusions (a key claim in our work), we conduct a systematic analysis of model performance across different occlusion severity levels. Following established occlusion assessment protocols in the 3DPW dataset, we categorize test sequences into three occlusion levels based on the percentage of visible body joints: Mild (80–100% joints visible), Moderate (50–79% joints visible), and Heavy (<50% joints visible). This categorization is performed automatically using 2D pose detection confidence scores from OpenPose [55], where joints with detection confidence below 0.3 are classified as occluded.

Table 5. Performance breakdown across different occlusion severity levels on 3DPW test set. Our method shows superior robustness, particularly in moderate and heavy occlusion scenarios where temporal–topological reasoning is most critical.

Method	MPJPE (mm) ↓			PA-MPJPE (mm) ↓
	Mild	Moderate	Heavy	Mild	Moderate	Heavy
Fastmetro	69.2	75.8	86.3	42.1	45.7	52.4
GLoT	75.1	82.4	94.7	47.3	51.2	59.8
4DHumans	65.8	71.3	81.5	41.9	45.6	52.1
Ours	68.1	73.7	79.8	41.2	44.3	49.6
Improvement over 4DHumans:
Ours	−2.3	−4.8	−7.2	+0.7	+1.3	−2.5

Note: ↓ indicates lower is better; Bold indicates best performance for each occlusion level; Italics in row headers indicate descriptive labels.

presents a detailed breakdown of reconstruction accuracy across these occlusion levels, comparing our ProGraph method with state-of-the-art approaches. Our method demonstrates particularly strong performance in moderate and heavy occlusion scenarios, achieving 4.8 mm and 7.2 mm improvements in MPJPE over 4DHumans, respectively. This validates our framework’s ability to leverage temporal probability distributions and topological priors for reconstructing occluded body regions.

Analysis of Occlusion Robustness. The results reveal several key insights. First, all methods perform reasonably well under mild occlusion conditions, with our approach achieving competitive but not dramatically superior results compared to 4DHumans. This is expected, as temporal reasoning provides limited additional benefit when most body parts are visible. Second, our method begins to demonstrate significant advantages (4.8 mm MPJPE improvement) in moderate occlusion scenarios (50–79% visibility), leveraging TPDist’s ability to propagate information from unoccluded frames to occluded regions through temporal probability alignment. Third, the performance gap widens to 7.2 mm MPJPE improvement under heavy occlusion (<50% visibility), demonstrating that our graph topological modeling effectively captures structural priors that constrain reconstruction even when direct visual evidence is severely limited. The hierarchical loss function (HHLoss) plays a crucial role in these challenging cases by enforcing part-level consistency across the temporal sequence.

Notably, our method maintains relatively stable performance degradation as occlusion severity increases (68.1 mm → 73.7 mm → 79.8 mm MPJPE), while competing methods exhibit steeper degradation curves. This suggests that our topological-probabilistic framework provides genuine robustness rather than merely optimizing for average-case performance.

Figure 4 illustrates the comparative effects of reconstructing different human motion sequences in two scenarios with missing or unclear body parts. Our ProGraph framework addresses these challenges by learning the probability distribution of each body part across the entire motion sequence.

4. Discussion

In this part, we discuss the significance of our study and its implications for the field of 3D human pose and shape reconstruction. Our ProGraph framework demonstrates the effectiveness of combining temporally alignable probability distributions with graph topological modeling, achieving state-of-the-art performance on challenging datasets.

The key strengths of our approach include: (1) a novel integration of probabilistic modeling with graph topology for human mesh reconstruction, (2) the ability to handle occlusions and motion blur through temporal consistency, and (3) a hierarchical loss function that enables accurate part-wise reconstruction.

4.1. Real-World Applications: Software-Defined Internet of Vehicles (SD-IoV)

An important emerging application domain for our robust 3D human mesh reconstruction framework is the Software-Defined Internet of Vehicles (SD-IoV), an intelligent transportation paradigm where vehicles communicate and process multimedia data over dynamically managed networks [56,57]. SD-IoV systems face inherent challenges in multimedia streaming due to packet loss, variable bandwidth, network latency, and motion-induced blur from vehicle movement. Our method’s ability to handle occluded and degraded video frames aligns directly with these practical challenges.

Specific SD-IoV Use Cases:

• Pedestrian Behavior Analysis for Autonomous Vehicles. Vehicle-mounted cameras must track pedestrian poses in real-time under adverse conditions including occlusions by other vehicles, motion blur from vehicle speed, and intermittent visibility due to network packet loss during video streaming. Our temporal–topological framework can reconstruct pedestrian body poses even when frames are missing or corrupted, enabling more reliable collision avoidance and trajectory prediction systems.
• Driver Activity Monitoring in Connected Vehicles. In-cabin cameras can monitor driver behavior (e.g., detecting distracted driving, fatigue) by analyzing 3D pose. However, driver body parts are frequently occluded by steering wheels, seatbelts, and cabin structures. Our GTM-TPDist architecture leverages body topology priors to infer occluded limb positions, while HHLoss ensures anatomically plausible pose estimates crucial for safety-critical driver state assessment.
• Vehicle-to-Everything (V2X) Video Communication. Video compression artifacts and dropped frames are common when vehicles share camera footage over bandwidth-constrained V2X networks. Our probabilistic temporal modeling can “fill in” missing information from degraded video streams, thereby maintaining continuous 3D human tracking for traffic monitoring and smart city applications.

Technical Synergies with SD-IoV Requirements. Our framework offers several technical advantages aligned with SD-IoV constraints. (1) Temporal coherence: TPDist’s sequence-level modeling tolerates the intermittent frame loss common in wireless vehicular networks. (2) Computational efficiency: As demonstrated in

Table 6. Computational efficiency comparison with 4DHumans. Our method achieves better accuracy with substantially fewer parameters and FLOPs.

Method	FLOPs ↓	Params ↓	3DPW		Human3.6M
			MPVE ↓	PA-MPJPE ↓	PA-MPJPE ↓
4DHumans	122,590.2 M	670.2 M	85.22	44.50	33.6
Ours	42,737.4 M	291.1 M	82.12	43.82	33.1

Note: ↓ indicates lower is better; Bold indicates best performance; Comma separators used for numbers with five or more digits.

, our method has reduced FLOPs (65.1% reduction vs. 4DHumans), making it more suitable for edge computing deployment on resource-constrained vehicle processors. (3) Occlusion robustness: Our occlusion analysis (Table 5) shows strong performance in heavy occlusion, which is critical for crowded urban traffic scenarios.

Future Research Directions in SD-IoV Contexts. Deploying our framework in SD-IoV systems opens several research avenues: (1) adaptive quality–bandwidth tradeoffs, dynamically adjusting reconstruction granularity based on available network bandwidth; (2) multi-vehicle collaborative perception, fusing 3D pose estimates from multiple vehicle viewpoints to compensate for individual camera occlusions; and (3) real-time optimization, further reducing latency through model quantization and hardware acceleration for safety-critical autonomous driving applications. These directions represent promising extensions of our work that leverage the unique characteristics of vehicular networks.

4.2. Limitations

While our ProGraph framework demonstrates strong performance across standard benchmarks, we acknowledge several limitations that warrant discussion and motivate future research directions:

1. Computational Complexity for Real-Time SD-IoV Applications. The temporal transformer and graph convolution operations introduce computational overhead that may limit deployment in latency-critical applications. Our current implementation achieves approximately 150 ms per frame on an NVIDIA A100 GPU, which exceeds the <100 ms latency requirement for real-time autonomous driving perception systems. For SD-IoV applications requiring edge deployment on resource-constrained vehicle processors (e.g., NVIDIA Jetson), further optimization through model quantization, knowledge distillation, or architectural simplification would be necessary. The transformer’s quadratic attention complexity with respect to sequence length also poses scalability challenges for processing longer video sequences.

2. Generalization to Non-SMPL Body Models. Our framework is inherently tied to the SMPL body model’s topology (6890 vertices, 24 joints). Extending it to alternative parametric models such as SMPL-X (with hands and face), STAR, or GHUM would require retraining the GTM module with new adjacency matrices and potentially redesigning the hierarchical loss structure. Furthermore, the SMPL model assumes standard adult body proportions, leading to suboptimal performance for several classes:

• Children and adolescents with different limb-to-torso ratios.
• Individuals with atypical body shapes (e.g., obesity, muscular builds).
• People with physical disabilities or prosthetic limbs.
• Non-human subjects (animals, robots) that may benefit from similar pose estimation.

3. Multi-Person Occlusion Scenarios. Our current framework processes single individuals and struggles in crowded scenes with complex inter-person occlusions. Specific failure modes include:

• Identity confusion: When two people overlap significantly (>50% mutual occlusion), the model may incorrectly associate limbs between individuals, producing chimeric reconstructions.
• Depth ambiguity: In scenes with multiple people at similar depths, the model cannot reliably determine which body parts belong to which person without explicit tracking.
• Prolonged occlusion: When a person is fully occluded for more than ten consecutive frames, temporal probability propagation becomes unreliable and the model may “hallucinate” implausible poses.

4. Sensitivity to Input Quality. Performance degrades significantly under extreme input degradation:

• Very low resolution (<128 × 128 pixels): Insufficient visual detail for reliable feature extraction.
• Severe motion blur (exposure >100ms): Loss of edge information critical for body part segmentation.
• Extreme lighting conditions: Overexposure or underexposure causing loss of texture information.

5. Training Data Bias. Our model is trained primarily on 3DPW and Human3.6M datasets, which have limited diversity in terms of certain aspects:

• Cultural pose variations (e.g., traditional dances, martial arts).
• Occupational activities (e.g., construction workers, athletes in specialized sports).
• Clothing diversity (loose garments, costumes that significantly alter body silhouette).

These limitations represent important directions for future work, including multi-person tracking integration, domain adaptation techniques, and model compression for edge deployment.

5. Conclusions

This paper presents a novel framework for robust 3D human mesh reconstruction from monocular video that addresses occlusion and blur challenges through temporal probability-guided graph topology learning. By constructing probabilistic distributions over body topology across temporal frames, our approach achieves significant improvements in 3D pose regression accuracy. Experimental results demonstrate superior performance compared to state-of-the-art methods on the 3DPW and Human3.6M benchmarks, with particularly strong robustness under challenging visual conditions. This work extends our preliminary arXiv preprint [28] with comprehensive implementation details, computational efficiency analysis, and additional experimental validation. The proposed framework’s robustness enables diverse applications requiring accurate 3D pose estimation in unconstrained real-world settings.

Future research directions include enhancing computational efficiency, extending the framework to multi-person scenarios with complex interactions, and addressing extreme occlusion cases with prolonged temporal coverage.

6. Enhancements to the Preprint Version

Relationship to arXiv Preprint: This manuscript substantially extends our preliminary work published as an arXiv preprint [28]. While the core methodology (GTM, TPDist, HHLoss) remains consistent to maintain scientific integrity, this journal submission provides significant additional contributions not present in the preprint version, specifically:

• Comprehensive implementation details enabling full reproducibility (Section 6.1).
• Computational efficiency analysis with detailed comparison to state-of-the-art methods (Section 6.2).
• Extended experimental validation and qualitative analysis (Section 6.3).
• Refined presentation with improved clarity and additional technical specifications.

These enhancements represent substantial additional work beyond the preliminary preprint and provide the readership with a complete, reproducible, and thoroughly validated contribution.

6.1. Implementation Specifications

This paper provides comprehensive implementation details for reproducibility. Our model parameters are initialized randomly, processing video sequences of 8 frames with a batch size of 16 sequences. Input frames are cropped to include the human subject and resized to $224\times 224$ pixels.

The feature extraction backbone utilizes HRNet-W64 [58] pretrained on ImageNet. The U-Net architecture incorporates cross-attention and 3D convolution in each encoder (downblock) and decoder (upblock) layer. Cross-attention facilitates effective information transfer from CLIP [59], while 3D convolution distributes data along the temporal dimension for alignment with the spatiotemporal midblock layer designed for hidden feature processing.

The prior model builds upon statistical frameworks [36], encoding high-confidence generalizable associations between body parts for robust generalization across reconstruction tasks. Our graph structure decouples and extracts this implicit association information.

For transformer initialization, we extend the GTM and temporal layer concepts. GTM first strengthens structural understanding among flattened joint and vertex tokens, then multi-head attention jointly processes camera and feature tokens. Inspired by PoseFormerV2 [60], we enhance information propagation between adjacent sequence tokens for improved motion stability.

Training employs the AdamW optimizer for 100 epochs with initial learning rate $1\times {10}^{-5}$, reduced by a factor of 10 after ten epochs. All experiments were run on PyTorch 2.0 with an Intel Xeon Platinum 8358 CPU @ 2.60 GHz and two NVIDIA A100 GPUs.

6.2. Computational Efficiency Analysis

We provide detailed computational comparison with a prior state-of-the-art method, namely, 4DHumans [22]. 4DHumans builds on PHALP for temporal person association, while our TPDist aligns probability distributions of 3D body geometry across consecutive frames through temporal guidance.

As shown in Table 6, our method outperforms 4DHumans across the MPVPE and PA-MPJPE metrics while significantly reducing computational requirements. Compared to 4DHumans, our approach reduces FLOPs by 65.1% and parameters by 56.6%, demonstrating superior efficiency alongside improved accuracy.

6.3. Additional Qualitative Analysis

In addition to quantitative metrics, we provide extensive qualitative analysis demonstrating our method’s robustness in challenging real-world conditions. This analysis extends our preliminary arXiv preprint by including additional failure case analysis, cross-dataset generalization evaluation, and detailed visual comparisons under extreme conditions.

Occlusion Handling. As illustrated in Figure 4 and Figure 5, our method successfully reconstructs occluded body parts by leveraging temporal consistency. When a subject’s legs are occluded by foreground objects in frames 15–20 of the “downtown” sequence in 3DPW, competing methods (Fastmetro, GLoT) produce anatomically implausible limb configurations. In contrast, our GTM-TPDist framework maintains plausible leg poses by propagating information from unoccluded frames through probabilistic temporal modeling, demonstrating the effectiveness of our topological-probabilistic approach.

Motion Blur Robustness. Our method exhibits superior stability in the presence of significant motion blur, which is common in fast-moving sequences. The “courtyard backpack” sequence contains frames with substantial blur during rapid turning motions. Baseline methods produce jittery reconstructions with frame-to-frame discontinuities, while our temporal probability distribution mechanism (TPDist) enforces smooth transitions, resulting in temporally coherent pose sequences even under degraded visual quality.

Comparison with 4DHumans. Figure 6 provides detailed visual comparison with 4DHumans. The red circles highlight critical differences in foot placement accuracy. In scenarios with ground-level occlusion, 4DHumans occasionally produces floating feet or ground penetration artifacts, whereas our hierarchical loss (HHLoss) enforces anatomically consistent lower-body configurations. This demonstrates that our vertex-level modeling provides finer geometric control compared to parameter-based approaches.

Cross-Dataset Generalization. We evaluated models trained on Human3.6M and tested on 3DPW without fine-tuning to assess generalization. Our method maintains reasonable reconstruction quality (MPJPE degradation of only 8.2 mm), while Fastmetro degrades by 14.7 mm. This suggests that our graph-topological priors capture domain-invariant body structure, enabling better cross-dataset transfer.

Failure Cases. We acknowledge limitations in extreme scenarios. First, when multiple people heavily overlap (more than 70% mutual occlusion for >10 consecutive frames), our single-person framework struggles to disentangle identities. Second, for extremely low-resolution inputs (<128 × 128 pixels), insufficient visual detail causes both our method and the baselines to fail. Third, non-standard body proportions (e.g., children, exceptionally tall individuals) sometimes lead to suboptimal fitting due to SMPL template biases. These cases represent promising directions for future work that could potentially be addressed through multi-person tracking integration, super-resolution preprocessing, and parametric model diversification.