Review Reports - Harmonizing Scale for Intelligent Sensors: Resolution-Conditioned Adaptation of Vision Foundation Models for Crowd Counting

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a crowd counting and localization method based on frozen DINOv3 features and Scale Coordination Adapter (SHA).
1. It is suggested to rewrite the contribution section to highlight the contribution of "introducing the foundation model into the selective inheritance paradigm" and weaken the original claim to the SHA module itself.
2. There is a lack of discussion on resolution related methods such as HMT and PARGT in the past two years, as well as on Transformer based crowd counting methods such as CCTrans and CrowdFormer.
3. Need to improve the quality of the image.
4.Most of the methods compared in the article are from 2019 to 2022 (STEERER is from 2023). In the past two years (2024-2026), there have been many methods in the field of crowd counting based on Transformers, diffusion models, or more advanced foundation models (such as CLIP-Counting, CrowdDiff, or direct fine-tuning based on DINOv2).

5.Although the article reports that the number of trainable parameters (28.4M) is lower than that of STEERER-HRNet (65.8M), the forward computational complexity of DINOv3-ViT-L/14 itself is much greater than that of HRNet-W48. Although freezing the backbone does not update parameters, FLOPs and latency during inference are crucial, especially for the application scenario of "intelligent monitoring sensors".

6.How does the receptive field of the soft mask generator in SHA affect the spatial smoothness of the attention map when using three convolutional layers? Lack of analysis.

Author Response

Journal Sensors (ISSN 1424-8220)

Manuscript ID: sensors-4235584

Type: Article

Title: Harmonizing Scale for Intelligent Sensor: Resolution-Conditioned Adaptation of Vision Foundation Models for Crowd Counting

Authors:

Huan Xu, Zhiheng Chen, Sirou Shen, Haibin You, Chih-Cheng Chen*

Section: Intelligent Sensors

Dear Editor,

Thank you very much for your letter and for the comments by the reviewers. These comments are very valuable and helpful for our paper.

We carefully revised the manuscript according to the reviewers' suggestions. All textual changes in the revised manuscript are highlighted in red italics.

Please contact us if any further questions remain.

Sincerely yours,

Prof. Chih-Cheng Chen

Response to the comments of reviewers:

Reviewer 4:

Comment 1. The novelty over existing hybrid feature adaptation and optimization-driven learning frameworks is not sufficiently clarified.

Response. We agree and clarified the paper positioning in the revised related-work and introduction sections. The revised manuscript now states more explicitly that our goal is not to design a new optimization-driven convergence mechanism, but to calibrate frozen foundation features so that they can be used inside the selective inheritance pipeline.

Revised text in manuscript: Our primary contribution is to show how a frozen vision foundation model can be introduced into the selective inheritance paradigm through resolution-conditioned adaptation, rather than to redefine selective inheritance itself.

In contrast to hybrid optimization-driven feature selection frameworks, our goal is not to design a new convergence mechanism, but to calibrate frozen foundation features so that they can operate within the selective inheritance pipeline.

Comment 2. There is no theoretical analysis explaining convergence behavior, feature selection optimality, or stability under scale variation.

Response. We appreciate this suggestion. In the revised manuscript, we intentionally toned down any wording that could be read as claiming theoretical optimality, and instead positioned SHA and the soft-mask generator as empirically validated practical design choices. Since the present work is empirical rather than theoretical, we do not claim convergence guarantees or scale-selection optimality. We revised the manuscript accordingly and believe this more accurately reflects the scope of the current paper.

Revised text in manuscript: Using three stacked 3 x 3 convolutions gives the SMG a moderate local effective receptive field, which empirically produces smoother spatial masks than a shallower design while still preserving fine crowd boundaries; we therefore treat this as a practical design choice rather than a theoretically optimal setting. Since the present work is empirical rather than theoretical, we do not claim convergence guarantees or scale-selection optimality.

Comment 3. PWSP lacks justification for optimality and comparison with alternative scale-selection strategies.

Response. We agree that the original wording could be interpreted too strongly. In the revision, we clarified the training/inference role of PWSP and avoided any implication of theoretical optimality. Since the present work is empirical rather than theoretical, we do not claim convergence guarantees or scale-selection optimality. The manuscript now states clearly that winner masks are training-only supervision signals, while the actual inference pipeline remains a single forward pass. We also emphasize the adaptive weighting role of the soft-mask pathway as an empirical design rather than a universally optimal scale-selection rule.

Revised text in manuscript: Importantly, these winner masks are constructed only during training from prediction errors against the ground-truth density maps. During inference, no ground truth or additional scale-selection step is required.

Using three stacked 3 x 3 convolutions gives the SMG a moderate local effective receptive field, which empirically produces smoother spatial masks than a shallower design while still preserving fine crowd boundaries; we therefore treat this as a practical design choice rather than a theoretically optimal setting. Since the present work is empirical rather than theoretical, we do not claim convergence guarantees or scale-selection optimality.

Comment 4. The datasets do not fully reflect real-world deployment challenges such as sensor noise, streaming data, and dynamic environments.

Response. We agree and have added a dedicated limitations and deployment-scope discussion. This new paragraph explicitly states that severe perspective distortion, domain shift, and sensor noise can weaken the progressive inheritance assumption, and that the current study should not be over-generalized to all real-world streaming scenarios.

Comment 5. The experimental comparison includes classical and selective inheritance baselines such as STEERER, CAN, and DM-Count but lacks comparison with optimization-based learning, reinforcement learning-based approaches, and adaptive feature selection frameworks.

Response. We respectfully disagree that these categories are necessary direct baselines for the present manuscript. The focus of this paper is crowd counting and localization under the selective inheritance setting, rather than generic optimization or policy-learning formulations. For this reason, we prioritize comparisons with representative crowd counting methods that use comparable supervision, outputs, and benchmark protocols. Many optimization-based or reinforcement-learning-based feature selection frameworks are designed for different objectives and assumptions, and they are not standard baselines in current crowd-counting benchmarks. We therefore believe that the current comparison set is the more appropriate and fairer one for evaluating the contribution of frozen foundation-feature adaptation in this task.

Comment 6. Although the paper uses standard datasets (ShanghaiTech, UCF-QNRF, JHU-Crowd++), these datasets do not fully reflect real-world deployment challenges such as sensor noise, streaming data, and dynamic environments.

Response. We appreciate this perspective, but we respectfully note that ShanghaiTech, UCF-QNRF, and JHU-Crowd++ remain the standard public benchmarks for evaluating crowd counting accuracy and generalization, and they are also the datasets used by the representative methods compared in our paper. Replacing them in the current revision with custom streaming or sensor-noise settings would reduce comparability rather than improve fairness. We therefore retain the standard benchmark protocol for the main evaluation and treat sensor noise, streaming input, and dynamic deployment conditions as important but separate extensions beyond the scope of the present study.

Comment 7. The paper claims efficiency, but does not discuss inference latency, memory footprint, or deployment feasibility on edge devices.

Response. We agree and addressed this directly by adding total parameters, average latency, FPS, and peak memory. These results show that the current DINOv3-L instantiation is not an ultra-light edge model. We therefore revised the manuscript to discuss trainable-parameter efficiency and deployment scope more carefully.

Revised text in manuscript: The new efficiency subsection reports deployment-oriented metrics on the same hardware. Compared with STEERER-HRNet, D3-CalibCount retains a smaller trainable part but incurs a larger total model size and higher inference cost because the frozen DINOv3-L backbone dominates the forward pass. We therefore revise our claim from general 'efficiency' to trainable-parameter efficiency and discuss the resulting accuracy-cost trade-off more explicitly.

The revised experiments also show a clear accuracy-cost trade-off: frozen foundation-feature adaptation improves counting accuracy and preserves training stability, but its current DINOv3-L instantiation is better suited to GPU-backed intelligent sensor nodes than to strict edge-device deployment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper proposes D3-CalibCount, a framework that adapts frozen DINOv3 features for crowd counting using a Scale Harmonization Adapter (SHA). The method builds on STEERER, replacing CNN backbones with a frozen ViT and adding lightweight adaptation modules. Experiments on three benchmarks show modest improvements over baselines.

Major Strengths

The idea of leveraging frozen foundation models for dense prediction tasks is timely and relevant.
The parameter-efficient design (only 28.4M trainable parameters) is a practical advantage for deployment.
The SHA module is reasonably well-motivated, and the ablation studies validate its components.

Major Concerns Requiring Revision

Missing computational efficiency analysis. The paper claims parameter efficiency but reports only trainable parameters. The frozen DINOv3 backbone itself contains ~304M parameters, making total model size substantially larger than HRNet baselines (~65M). The authors must report inference-time metrics: FLOPs, frames per second (FPS), and memory usage. Without these, the efficiency claims are incomplete and potentially misleading.
Modest performance gains. Improvements over STEERER-HRNet are small: 5.0% MAE reduction on SHHA, 4.2% on UCF-QNRF. Given the much larger backbone, the paper should discuss the trade-off between accuracy gain and computational cost more explicitly. Are these gains statistically significant? The authors should report standard deviations or confidence intervals.
Missing comparisons. The paper lacks comparisons to:

Other frozen foundation models (DINOv2, CLIP, MAE, SAM) to isolate what DINOv3 specifically contributes.
Recent 2024-2025 crowd counting methods to establish state-of-the-art positioning.
Fine-tuned DINOv3 (not just frozen) with comparable parameter budgets.

Unclear training protocol. The Masked Selection and Inheritance Loss uses per-pixel winner selection across resolutions. It is unclear whether this requires ground-truth density maps during training (standard) or during inference (problematic). The description in Section 3.4.1 needs clarification. Also, training for 2,000 epochs on small datasets like ShanghaiTech (300 training images) raises overfitting concerns. The authors should provide validation curves or early stopping criteria.
Incomplete localization results. Section 4.2.4 promises precision and recall metrics but only reports partial F1-scores. Full localization results should be provided.

Minor Issues

The paper would benefit from a discussion of failure cases or limitations.
The connection to Sensors journal (sensor systems) is not clearly established. The authors should clarify relevance. Now there only a couple of words in the abstract on the subject.

Required for revision:

FLOPs, FPS, and memory comparisons with baselines:

Comparisons to other foundation models (DINOv2, CLIP, etc.);

Clarification of PWSP training vs. inference;

Validation curves or early stopping justification;

Complete localization metrics.

Author Response

Journal Sensors (ISSN 1424-8220)

Manuscript ID: sensors-4235584

Type: Article

Title: Harmonizing Scale for Intelligent Sensor: Resolution-Conditioned Adaptation of Vision Foundation Models for Crowd Counting

Authors:

Huan Xu, Zhiheng Chen, Sirou Shen, Haibin You, Chih-Cheng Chen*

Section: Intelligent Sensors

Dear Editor,

Thank you very much for your letter and for the comments by the reviewers. These comments are very valuable and helpful for our paper.

We carefully revised the manuscript according to the reviewers' suggestions. All textual changes in the revised manuscript are highlighted in red italics.

Please contact us if any further questions remain.

Sincerely yours,

Prof. Chih-Cheng Chen

Response to the comments of reviewers:

Reviewer 2:

Comment 1. Missing computational efficiency analysis.

Response. We agree and have added a dedicated efficiency and deployment subsection. The revised manuscript now reports total parameters, average latency, FPS, and peak memory on the same hardware. These results make the trade-off explicit: our method uses a smaller trainable adaptation part, but the frozen DINOv3-L backbone increases total model size and inference cost relative to STEERER-HRNet.

Accordingly, we revised the wording throughout the paper from general efficiency to trainable-parameter efficiency.

Comment 2. Modest performance gains.

Response. We agree that the gains should be presented in a measured way. In the revision, we explicitly discuss the accuracy-cost trade-off and avoid overclaiming. We now describe the improvements as consistent but modest gains over a strong selective inheritance baseline, rather than as a blanket efficiency advantage.

Revised text in manuscript: The revised experiments also show a clear accuracy-cost trade-off: frozen foundation-feature adaptation improves counting accuracy and preserves training stability, but its current DINOv3-L instantiation is better suited to GPU-backed intelligent sensor nodes than to strict edge-device deployment.

Comment 3. Missing comparisons to other foundation models and recent methods.

Response. We agree that a broader foundation-model benchmark is important. In this revision, we addressed this in three ways: first, we expanded the related-work discussion to include representative recent Transformer-based and pre-trained-representation-based crowd counting methods; second, we added a direct DINOv2 vs. DINOv3 comparison under the same SHA-based adaptation pipeline; third, we retained the frozen vs. fine-tuned DINOv3 ablation because it directly addresses whether the gain comes from the adaptation setting or from full backbone tuning.

Revised text in manuscript: Recent transformer-based or pre-trained-representation-based counting methods, such as CCTrans, CrowdFormer, and CrowdCLIP, further highlight the value of stronger global representations, but they do not study how frozen foundation features interact with selective inheritance across resolutions.

A direct comparison between DINOv2 and DINOv3 under the same SHA-based adaptation pipeline shows that DINOv3 performs better on both UCF-QNRF and JHU-Crowd++, supporting our choice of backbone while keeping the rest of the framework fixed.

A broader comparison against additional foundation backbones beyond the DINOv2 baseline, such as vision-language models, is also an important next step.

Comment 4. Unclear training protocol, including PWSP and the long training schedule.

Response. We have clarified both points in the revised manuscript. PWSP winner masks are constructed only during training by comparing multi-resolution prediction errors against the ground-truth density maps; no ground truth is required during inference. We also added a short explanation that, despite the long nominal schedule, only the lightweight SHA modules and the counting head are optimized, while the large DINOv3 backbone remains frozen. This substantially reduces the overfitting risk compared with full-backbone fine-tuning.

Revised text in manuscript: Importantly, these winner masks are constructed only during training from prediction errors against the ground-truth density maps. During inference, no ground truth or additional scale-selection step is required.

Although the nominal number of epochs is large, only the SHA modules and the lightweight counting head are optimized, while the DINOv3 backbone remains frozen; this substantially reduces the overfitting risk compared with full-backbone fine-tuning and follows a stable training recipe inherited from the selective inheritance framework.

Comment 5. Incomplete localization results.

Response. We understand this concern. To remain consistent with common crowd localization practice, the revised manuscript now explicitly states that F1-score across multiple distance thresholds is the primary comparison metric, while representative precision and recall values are additionally reported at the standard threshold sigma_l = 2 for interpretation. In other words, we clarified the protocol and avoided implying that the paper provides a full precision-recall benchmark table against all prior works.

Revised text in manuscript: For localization, following common crowd localization practice, we use F1-score across multiple distance thresholds as the primary comparison metric and additionally report representative precision and recall at a standard threshold to aid interpretation.

Following common crowd localization practice, we evaluate on ShanghaiTech Part A primarily with F1-score under distance thresholds of sigma_l in {1, 2, 3} pixels and report representative precision/recall at sigma_l = 2.

Comment 6. The paper would benefit from a discussion of failure cases or limitations.

Response. We agree and have added a dedicated limitations and deployment-scope paragraph. This new discussion acknowledges the higher inference cost of the current DINOv3-L instantiation and the remaining challenges under severe perspective distortion, domain shift, and sensor noise.

Comment 7. The connection to Sensors journal is not clearly established.

Response. We strengthened this connection by revising the abstract and conclusion so that the deployment scope is stated more explicitly. We now clarify that the target scenario is camera-based intelligent surveillance sensing, where crowd density estimation is performed from visual sensor inputs at GPU-backed monitoring nodes.

Revised text in manuscript: From a deployment perspective, the method reduces the trainable portion of the network, while the frozen DINOv3-L backbone still introduces a higher inference cost than lighter CNN baselines. The target application scenario is camera-based intelligent surveillance sensing, where crowd density estimation is performed from visual sensor inputs at GPU-backed monitoring nodes.

The revised experiments also show a clear accuracy-cost trade-off: frozen foundation-feature adaptation improves counting accuracy and preserves training stability, but its current DINOv3-L instantiation is better suited to camera-based intelligent surveillance sensing at GPU-backed monitoring nodes than to strict edge-device deployment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The actual scientific contribution of the paper should be defined more precisely. The authors should clearly distinguish the components adopted from STEERER from those that are genuinely novel.
The authors repeatedly claim that frozen DINOv3 features are not naturally aligned with density estimation and localization, yet they do not provide formal evidence to support this claim.
The current set of comparative methods is useful, but insufficient. Comparisons with stronger and more recent methods based on transformers, foundation models, and related approaches should be added.
A more complete analysis of the computational and deployment costs should be included.
The paper claims to contribute to both crowd counting and localization, yet the localization section is relatively modest.
The ablation studies are appropriate, but still too basic. The authors should show the impact of the choice of DINOv3 layers, the number of hierarchy levels, the value of λ in the loss function, the SMG structure, the channel dimensions in SHA, and different variants of the counting head.
It would be worthwhile to include an analysis of failure cases and the limitations of the proposed method.

Author Response

Journal Sensors (ISSN 1424-8220)

Manuscript ID: sensors-4235584

Type: Article

Title: Harmonizing Scale for Intelligent Sensor: Resolution-Conditioned Adaptation of Vision Foundation Models for Crowd Counting

Authors:

Huan Xu, Zhiheng Chen, Sirou Shen, Haibin You, Chih-Cheng Chen*

Section: Intelligent Sensors

Dear Editor,

Thank you very much for your letter and for the comments by the reviewers. These comments are very valuable and helpful for our paper.

We carefully revised the manuscript according to the reviewers' suggestions. All textual changes in the revised manuscript are highlighted in red italics.

Please contact us if any further questions remain.

Sincerely yours,

Prof. Chih-Cheng Chen

Response to the comments of reviewers:

Reviewer 3:

Comment 1. The actual scientific contribution of the paper should be defined more precisely.

Response. We agree and have rewritten the introduction and contribution section to separate inherited components from genuinely new ones. The revised text now makes it explicit that the main contribution is introducing a frozen foundation model into the selective inheritance paradigm through resolution-conditioned adaptation, rather than redefining the selective inheritance framework itself.

Revised text in manuscript: Our primary contribution is to show how a frozen vision foundation model can be introduced into the selective inheritance paradigm through resolution-conditioned adaptation, rather than to redefine selective inheritance itself.

We introduce frozen foundation representations into the selective inheritance paradigm for crowd counting and show that resolution-conditioned calibration is needed to make them compatible with density estimation and cross-scale inheritance.

Comment 2. The authors repeatedly claim that frozen DINOv3 features are not naturally aligned with density estimation and localization, yet they do not provide formal evidence.

Response. We revised this point in two ways. First, we toned down the language so that the claim is presented as an empirical observation rather than as a formal theorem. Second, we added an explicit sentence in the backbone ablation discussion stating that the gap between DINOv3 without SHA and DINOv3 with SHA provides empirical evidence that strong frozen features alone are still not fully aligned with selective scale-aware counting.

Revised text in manuscript: At the same time, the gap between DINOv3 (w/o SHA) and DINOv3 (w/ SHA) provides empirical evidence that strong frozen foundation features alone are not fully aligned with selective scale-aware counting, and that resolution-conditioned calibration remains necessary.

Comment 3. The current set of comparative methods is useful, but insufficient.

Response. We agree that the manuscript needed clearer positioning with respect to recent Transformer-based and foundation-model-based directions. In the revision, we expanded the related-work discussion, added a direct DINOv2 vs. DINOv3 comparison under the same SHA-based adaptation pipeline, and included an explicit limitation statement that a broader comparison beyond the DINOv2 baseline remains important future work. At the same time, we avoided overclaiming exhaustive superiority beyond the comparisons actually reported.

Comment 4. A more complete analysis of the computational and deployment costs should be included.

Response. We agree and added a dedicated efficiency and deployment discussion with total parameters, average latency, FPS, and peak memory. This new subsection makes clear that the current model trades higher deployment cost for better accuracy, and we revised the manuscript wording accordingly.

Comment 5. The paper claims to contribute to both crowd counting and localization, yet the localization section is relatively modest.

Response. We agree and have adjusted the wording so that localization is presented as a supportive evaluation of density-map quality rather than as the primary contribution of the paper. We also clarified the metric protocol to match common reporting practice in crowd localization.

Revised text in manuscript: We therefore view localization here as a supportive evaluation of density-map quality rather than the main contribution of the paper.

Comment 6. The ablation studies are appropriate, but still too basic.

Response. We agree that a larger hyper-parameter sweep would be informative. For this revision, however, we intentionally focused the ablation on the factors most directly tied to the paper's main claim: backbone choice, adapter structure, and frozen-versus-fine-tuned training. We added this scope statement explicitly so that the paper does not overstate the breadth of the current ablation study.

Revised text in manuscript: Given the scope of the revision, we focus the ablation on backbone choice, adapter structure, and frozen-versus-fine-tuned training, which are the factors most directly tied to the paper's main claim.

Comment 7. It would be worthwhile to include an analysis of failure cases and the limitations of the proposed method.

Response. We agree and added a dedicated limitations and deployment-scope paragraph. This new text acknowledges the current model's higher inference cost and the remaining challenges under severe perspective distortion, domain shift, and sensor noise.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

The paper proposes D3-CalibCount with a Scale Harmonization Adapter (SHA) to adapt frozen DINOv3 features for scale-aware crowd counting. However, several critical issues related to novelty clarification, theoretical justification, experimental validation, and positioning within recent literature must be addressed before the manuscript can be considered for publication.

The novelty over existing hybrid feature adaptation and optimization-driven learning frameworks is not sufficiently clarified. Recent hybrid optimization-driven learning frameworks have explored structured adaptation strategies for improving convergence and feature selection (doi:10.1109/ACCESS.2021.3113350), which should be discussed to better position the contribution.
The proposed Scale Harmonization Adapter (Section 3.3) introduces dual-branch feature decomposition and soft-mask fusion (Eqs. (1)–(5)) . However, there is no theoretical analysis explaining:
- convergence behavior
- feature selection optimality
- stability under scale variation
- Provide theoretical insights or justification of why dual-path + soft-mask improves scale alignment.
The Patch-Winner Selection Principle (PWSP) (Eq. (7)) dynamically assigns scale responsibility. However:
- No justification is given for its optimality
- No comparison with alternative scale-selection strategies
Compare with attention-based or adaptive weighting mechanisms.
The method assumes that scale harmonization across resolutions is consistent via progressive inheritance (Section 3.3–3.4) . However, in real-world scenarios:
- severe perspective distortion
- domain shifts
may break this assumption.
The experimental comparison includes classical and selective inheritance baselines such as STEERER, CAN, and DM-Count but lacks comparison with:
- optimization-based learning
- reinforcement learning-based approaches
- adaptive feature selection frameworks
Although the paper uses standard datasets (ShanghaiTech, UCF-QNRF, JHU-Crowd++) , these datasets do not fully reflect real-world deployment challenges such as:
- sensor noise
- streaming data
- dynamic environments
The paper claims efficiency (28.4M parameters vs 65.8M), but does not discuss:
- inference latency
- memory footprint
- deployment feasibility on edge devices

Comments on the Quality of English Language

Language should be improved

Author Response

Comment 1. The novelty over existing hybrid feature adaptation and optimization-driven learning frameworks is not sufficiently clarified.

Revised text in manuscript: Our primary contribution is to show how a frozen vision
foundation model can be introduced into the selective inheritance paradigm through resolutionconditioned
adaptation, rather than to redefine selective inheritance itself.
In contrast to hybrid optimization-driven feature selection frameworks, our goal is not to design
a new convergence mechanism, but to calibrate frozen foundation features so that they can
operate within the selective inheritance pipeline.

Comment 2. There is no theoretical analysis explaining convergence behavior, feature selection
optimality, or stability under scale variation.

Response. We appreciate this suggestion. In the revised manuscript, we intentionally toned
down any wording that could be read as claiming theoretical optimality, and instead positioned
SHA and the soft-mask generator as empirically validated practical design choices. Since the
present work is empirical rather than theoretical, we do not claim convergence guarantees or
scale-selection optimality. We revised the manuscript accordingly and believe this more
accurately reflects the scope of the current paper.

Revised text in manuscript: Using three stacked 3 x 3 convolutions gives the SMG a moderate
local effective receptive field, which empirically produces smoother spatial masks than a
shallower design while still preserving fine crowd boundaries; we therefore treat this as a
practical design choice rather than a theoretically optimal setting. Since the present work is
empirical rather than theoretical, we do not claim convergence guarantees or scale-selection
optimality.

Comment 3. PWSP lacks justification for optimality and comparison with alternative scaleselection
strategies.

Response. We agree that the original wording could be interpreted too strongly. In the revision,
we clarified the training/inference role of PWSP and avoided any implication of theoretical
optimality. Since the present work is empirical rather than theoretical, we do not claim
convergence guarantees or scale-selection optimality. The manuscript now states clearly that
winner masks are training-only supervision signals, while the actual inference pipeline remains a
single forward pass. We also emphasize the adaptive weighting role of the soft-mask pathway as
an empirical design rather than a universally optimal scale-selection rule.

Revised text in manuscript: Importantly, these winner masks are constructed only during
training from prediction errors against the ground-truth density maps. During inference, no
ground truth or additional scale-selection step is required.
Using three stacked 3 x 3 convolutions gives the SMG a moderate local effective receptive field,
which empirically produces smoother spatial masks than a shallower design while still
preserving fine crowd boundaries; we therefore treat this as a practical design choice rather
than a theoretically optimal setting. Since the present work is empirical rather than theoretical,
we do not claim convergence guarantees or scale-selection optimality.

Comment 4. The datasets do not fully reflect real-world deployment challenges such as sensor
noise, streaming data, and dynamic environments.

Response. We agree and have added a dedicated limitations and deployment-scope discussion.
This new paragraph explicitly states that severe perspective distortion, domain shift, and sensor
noise can weaken the progressive inheritance assumption, and that the current study should not
be over-generalized to all real-world streaming scenarios.

Revised text in manuscript: The present model is better suited to GPU-backed intelligent
surveillance nodes than to ultra-light edge devices. Although freezing DINOv3 reduces the
trainable portion of the network, the large backbone still increases total parameters, latency,
and memory footprint during inference. In addition, severe perspective distortion, domain shift,
or sensor noise can weaken the consistency assumption behind progressive inheritance,
occasionally leading to underestimation in extremely dense distant regions or over-smoothing in
sparse foreground regions.

Comment 5. The experimental comparison includes classical and selective inheritance baselines
such as STEERER, CAN, and DM-Count but lacks comparison with optimization-based
learning, reinforcement learning-based approaches, and adaptive feature selection frameworks.

Response. We respectfully disagree that these categories are necessary direct baselines for the
present manuscript. The focus of this paper is crowd counting and localization under the
selective inheritance setting, rather than generic optimization or policy-learning formulations.
For this reason, we prioritize comparisons with representative crowd counting methods that use
comparable supervision, outputs, and benchmark protocols. Many optimization-based or
reinforcement-learning-based feature selection frameworks are designed for different objectives
and assumptions, and they are not standard baselines in current crowd-counting benchmarks. We
therefore believe that the current comparison set is the more appropriate and fairer one for
evaluating the contribution of frozen foundation-feature adaptation in this task.

Comment 6. Although the paper uses standard datasets (ShanghaiTech, UCF-QNRF, JHUCrowd++),
these datasets do not fully reflect real-world deployment challenges such as sensor
noise, streaming data, and dynamic environments.

Response. We appreciate this perspective, but we respectfully note that ShanghaiTech, UCFQNRF,
and JHU-Crowd++ remain the standard public benchmarks for evaluating crowd
counting accuracy and generalization, and they are also the datasets used by the representative
methods compared in our paper. Replacing them in the current revision with custom streaming or
sensor-noise settings would reduce comparability rather than improve fairness. We therefore
retain the standard benchmark protocol for the main evaluation and treat sensor noise, streaming
input, and dynamic deployment conditions as important but separate extensions beyond the
scope of the present study.

Comment 7. The paper claims efficiency, but does not discuss inference latency, memory
footprint, or deployment feasibility on edge devices.

Response. We agree and addressed this directly by adding total parameters, average latency,
FPS, and peak memory. These results show that the current DINOv3-L instantiation is not an
ultra-light edge model. We therefore revised the manuscript to discuss trainable-parameter
efficiency and deployment scope more carefully.

Revised text in manuscript: The new efficiency subsection reports deployment-oriented metrics
on the same hardware. Compared with STEERER-HRNet, D3-CalibCount retains a smaller
trainable part but incurs a larger total model size and higher inference cost because the frozen
DINOv3-L backbone dominates the forward pass. We therefore revise our claim from general
'efficiency' to trainable-parameter efficiency and discuss the resulting accuracy-cost trade-off
more explicitly.
The revised experiments also show a clear accuracy-cost trade-off: frozen foundation-feature
adaptation improves counting accuracy and preserves training stability, but its current DINOv3-
L instantiation is better suited to GPU-backed intelligent sensor nodes than to strict edge-device
deployment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Problems have been addressed according to comments.

Author Response

Journal Sensors (ISSN 1424-8220)

Manuscript ID: sensors-4235584

Type: Article

Title: Harmonizing Scale for Intelligent Sensor: Resolution-Conditioned Adaptation of Vision Foundation Models for Crowd Counting

Authors:

Huan Xu, Zhiheng Chen, Sirou Shen, Haibin You, Chih-Cheng Chen*

Section: Intelligent Sensors

Dear Editor,

Thank you very much for your letter and for the comments by the reviewers. These comments are very valuable and helpful for our paper.

We carefully revised the manuscript according to the reviewers' suggestions. All textual changes in the revised manuscript are highlighted in red italics.

Please contact us if any further questions remain.

Sincerely yours,

Prof. Chih-Cheng Chen

Response to the comments of reviewers:

Reviewer 1:

Comment 1. Problems have been addressed according to comments.

Response. We thank the reviewer for the positive assessment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1. Previusly suggested by me references

Gramformer (AAAI 2024) [Lin, Hui, et al. "Gramformer: Learning crowd counting via graph-modulated transformer." Proceedings of the AAAI conference on artificial intelligence. Vol. 38. No. 4. 2024]

and

CsViT (2024) [Liu, Shuang, et al. "Cross-scale Vision Transformer for crowd localization." Journal of King Saud University-Computer and Information Sciences 36.2 (2024): 101972]

must be cited and a comment on the relation with these articles is to be done.

2. FLOPS for the model are to be given.

Author Response

Journal Sensors (ISSN 1424-8220)

Manuscript ID: sensors-4235584

Type: Article

Title: Harmonizing Scale for Intelligent Sensor: Resolution-Conditioned Adaptation of Vision Foundation Models for Crowd Counting

Authors:

Huan Xu, Zhiheng Chen, Sirou Shen, Haibin You, Chih-Cheng Chen*

Section: Intelligent Sensors

Dear Editor,

Thank you very much for your letter and for the comments by the reviewers. These comments are very valuable and helpful for our paper.

We carefully revised the manuscript according to the reviewers' suggestions. All textual changes in the revised manuscript are highlighted in red italics.

Please contact us if any further questions remain.

Sincerely yours,

Prof. Chih-Cheng Chen

Response to the comments of reviewers:

Reviewer 2:

Comment 1. Previusly suggested by me references Gramformer (AAAI 2024) and CsViT (2024) must be cited and a comment on the relation with these articles is to be done.

Response. We have added both Gramformer and CsViT to the related-work section and clarified their relation to our method. In the revised manuscript, they are positioned as recent task-specific transformer models for crowd counting/localization, whereas our contribution is to adapt frozen DINO foundation features so that they can operate within the inherited selective inheritance pipeline.

Revised text in manuscript: More recent task-specific transformer variants, including Gramformer and CsViT, further strengthen long-range relation modeling or cross-scale interaction for counting and localization. Unlike these end-to-end transformer architectures, our goal is not to redesign the counting model from scratch, but to adapt frozen foundation features so that they work inside the selective inheritance framework.

Comment 2. FLOPS for the model are to be given.

Response. We have added the model complexity to the efficiency subsection and now report TFLOPs together with total parameters, average latency, FPS, and peak memory under the same inference setting. This makes the computational trade-off explicit and allows direct comparison with STEERER-HRNet.

Specifically, STEERER-HRNet requires 0.284 TFLOPs, whereas D3-CalibCount requires 1.484 TFLOPs under the same evaluation setting.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Thank you for the careful revision. The paper is clearly improved: the contribution is now easier to understand, the localization claim is more appropriately moderated, and the added discussion of efficiency and limitations makes the manuscript more balanced. Before publication, I would still ask for a final minor revision: please state more explicitly what is inherited from STEERER and what is genuinely new, keep the claims about DINOv3 strictly empirical, acknowledge that the comparison and ablation scope is still limited, and fix the remaining editorial issues in the text. Subject to these small corrections, I consider the manuscript acceptable.

Author Response

Journal Sensors (ISSN 1424-8220) Manuscript ID: sensors-4235584 Type: Article Title: Harmonizing Scale for Intelligent Sensor: Resolution-Conditioned Adaptation of Vision Foundation Models for Crowd Counting Authors: Huan Xu, Zhiheng Chen, Sirou Shen, Haibin You, Chih-Cheng Chen* Section: Intelligent Sensors Dear Editor, Thank you very much for your letter and for the comments by the reviewers. These comments are very valuable and helpful for our paper. We carefully revised the manuscript according to the reviewers' suggestions. All textual changes in the revised manuscript are highlighted in red italics. Please contact us if any further questions remain. Sincerely yours, Prof. Chih-Cheng Chen Response to the comments of reviewers: Reviewer 3: Comment 1. Thank you for the careful revision. The paper is clearly improved: the contribution is now easier to understand, the localization claim is more appropriately moderated, and the added discussion of efficiency and limitations makes the manuscript more balanced. Before publication, I would still ask for a final minor revision: please state more explicitly what is inherited from STEERER and what is genuinely new, keep the claims about DINOv3 strictly empirical, acknowledge that the comparison and ablation scope is still limited, and fix the remaining editorial issues in the text. Subject to these small corrections, I consider the manuscript acceptable. Response. We thank the reviewer for the positive assessment and for the precise final suggestions. We implemented these clarifications directly in the manuscript. The revised text now states more explicitly which parts are inherited from STEERER and which parts are newly introduced in this work, keeps the statements about DINOv3 strictly empirical, and explicitly acknowledges that the current comparison and ablation scope remains limited. We also corrected the remaining editorial issues in the text, including dataset/citation wording and duplicated method labels in the experimental section. Revised text in manuscript: Our primary contribution is to show how a frozen vision foundation model can be introduced into the selective inheritance paradigm through resolution-conditioned adaptation, rather than to redefine selective inheritance itself. Concretely, we inherit from STEERER the progressive multi-resolution learning framework, the PWSP/MSIL supervision strategy, and the coarse-to-fine inheritance order, while the frozen DINOv3 backbone integration, the SHA module, and the corresponding feature-calibration pipeline are newly introduced in this work. At the same time, the gap between DINOv3 (w/o SHA) and DINOv3 (w/ SHA) provides empirical evidence that strong frozen foundation features alone are not fully aligned with selective scale-aware counting, and that resolution-conditioned calibration remains necessary. The present comparison scope is still limited to standard crowd-counting benchmarks and a focused set of ablations on backbone choice, adapter design, and freezing strategy; broader comparisons against additional foundation backbones, robustness protocols, or alternative scale-selection mechanisms remain future work.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

The authors have improved the manuscript clarity and positioning; however, some issues remain only partially addressed.

Novelty clarification:
The distinction from existing hybrid feature adaptation and optimization-driven frameworks is still not sufficiently analyzed. A clearer comparative discussion is required.
Theoretical justification:
The manuscript lacks sufficient theoretical insight into:
- why dual-branch (CHP + FIP) improves scale alignment
- how soft-mask fusion contributes to stability
  Stronger conceptual or analytical justification is needed.
PWSP mechanism:
The Patch-Winner Selection Principle is still weakly justified.
No comparison with alternative approaches (e.g., attention-based or soft weighting) is provided, and its limitations are not discussed.
Experimental scope:
While standard datasets are acceptable, the work lacks evaluation under challenging real-world conditions such as:
- noise
- domain shift
- dynamic environments
Comparison with broader methods:
The justification for excluding optimization-based and reinforcement learning-based methods is reasonable.
Efficiency analysis:
Although latency, memory, and parameter analysis have been added, results indicate:
- high total parameters
- increased inference cost
  The claim of efficiency should be more carefully framed, and deployment limitations should be emphasized.
Clarity and presentation:
Some sections (especially SHA and PWSP explanations) remain dense and could benefit from clearer intuitive explanations.

Comments on the Quality of English Language

Language should be improved

Author Response

Journal Sensors (ISSN 1424-8220)

Manuscript ID: sensors-4235584

Type: Article

Title: Harmonizing Scale for Intelligent Sensor: Resolution-Conditioned Adaptation of Vision Foundation Models for Crowd Counting

Authors:

Huan Xu, Zhiheng Chen, Sirou Shen, Haibin You, Chih-Cheng Chen*

Section: Intelligent Sensors

Dear Editor,

Thank you very much for your letter and for the comments by the reviewers. These comments are very valuable and helpful for our paper.

We carefully revised the manuscript according to the reviewers' suggestions. All textual changes in the revised manuscript are highlighted in red italics.

Please contact us if any further questions remain.

Sincerely yours,

Prof. Chih-Cheng Chen

Response to the comments of reviewers:

Reviewer 4:

Comment 1. Novelty clarification: The distinction from existing hybrid feature adaptation and optimization-driven frameworks is still not sufficiently analyzed. A clearer comparative discussion is required.

Response. We strengthened this distinction in both the introduction and the related-work section. The revised manuscript now states explicitly that the inherited selective inheritance protocol remains unchanged, while the new contribution lies in introducing frozen DINOv3 features and the SHA-based calibration stage that makes those features compatible with that inherited protocol.

Comment 2. Theoretical justification: The manuscript lacks sufficient theoretical insight into why dual-branch (CHP + FIP) improves scale alignment and how soft-mask fusion contributes to stability. Stronger conceptual or analytical justification is needed.

Response. We did not add a formal proof, but we did make the conceptual justification more explicit in the manuscript. The revised text now explains the intended roles of CHP and FIP in plain terms, clarifies why separating immediate prediction from future inheritance is beneficial, and explains the soft mask as a continuous local gate that avoids abrupt branch switching. We also explicitly state that the present paper is empirical rather than theoretical and does not claim convergence guarantees or scale-selection optimality.

Revised text in manuscript: In intuitive terms, each SHA module decides which inherited evidence is already reliable enough to use at the current resolution and which evidence should be preserved and refined at the next finer resolution.

Intuitively, the dual-branch design separates two roles that a single branch would mix together: CHP emphasizes evidence that is already reliable at the current resolution, whereas FIP preserves complementary evidence that may become more discriminative only after it is propagated to finer resolutions. This separation reduces the conflict between immediate prediction and future inheritance, which is the main conceptual reason we observe better scale alignment in practice.

The soft mask serves as a local gate between CHP and FIP, so mixed-scale or ambiguous regions are not forced into a hard early decision and can retain complementary information for later resolutions. This continuous gating also avoids abrupt branch switching, which is the main intuitive reason it improves stability under mixed-scale regions. Since the present work is empirical rather than theoretical, we do not claim convergence guarantees or scale-selection optimality.

Comment 3. PWSP mechanism: The Patch-Winner Selection Principle is still weakly justified. No comparison with alternative approaches (e.g., attention-based or soft weighting) is provided, and its limitations are not discussed.

Response. We clarified the role and boundary of PWSP more directly. In the revised manuscript, PWSP is described as an inherited supervisory heuristic used only during training, not as a claim of optimal scale assignment. We also added its main limitation explicitly: because it is a hard winner-take-all signal, regions near scale boundaries can receive noisy supervision. Alternative soft weighting or attention-based scale-selection rules are possible, but they are beyond the scope of the present revision.

Revised text in manuscript: Intuitively, PWSP asks which resolution currently predicts each region best and uses that answer as a training-only supervisory cue.

We therefore treat PWSP as a simple and interpretable supervisory heuristic inherited from the selective inheritance framework, rather than as a claim of optimal scale assignment. A limitation of PWSP is that it is a hard winner-take-all assignment, so regions near scale boundaries can receive noisy supervision; alternative soft weighting or attention-based scale-selection rules are possible, but are beyond the scope of the present revision. For this reason, we use PWSP only as a training signal and rely on the continuous SHA features during inference.

Comment 4. Experimental scope: While standard datasets are acceptable, the work lacks evaluation under challenging real-world conditions such as noise, domain shift, and dynamic environments.

Response. We agree that these settings are important, but we do not believe that adding custom robustness protocols in this late revision would improve task-level comparability. The representative methods compared in the paper are evaluated on the same standard public benchmarks, so we retained the standard protocol for fair comparison and made the limitation explicit instead of overstating robustness beyond the evidence provided.

Revised text in manuscript: The present model is better suited to GPU-backed intelligent surveillance nodes than to ultra-light edge devices. Although freezing DINOv3 reduces the trainable portion of the network, the large backbone still increases total parameters, latency, and memory footprint during inference. In addition, severe perspective distortion, domain shift, or sensor noise can weaken the consistency assumption behind progressive inheritance, occasionally leading to underestimation in extremely dense distant regions or over-smoothing in sparse foreground regions. The present comparison scope is still limited to standard crowd-counting benchmarks and a focused set of ablations on backbone choice, adapter design, and freezing strategy; broader comparisons against additional foundation backbones, robustness protocols, or alternative scale-selection mechanisms remain future work.

Comment 5. Comparison with broader methods: The justification for excluding optimization-based and reinforcement learning-based methods is reasonable.

Response. We thank the reviewer for recognizing this point. We therefore keep the comparison protocol centered on representative crowd-counting methods with matched supervision and benchmark settings, which we believe is the fairest way to evaluate the contribution of frozen foundation-feature adaptation in this task.

Comment 6. Efficiency analysis: Although latency, memory, and parameter analysis have been added, results indicate high total parameters and increased inference cost. The claim of efficiency should be more carefully framed, and deployment limitations should be emphasized.

Response. We agree and tightened this framing further. The revised manuscript now reports TFLOPs in addition to total parameters, latency, FPS, and peak memory, and it consistently frames the advantage as trainable-parameter efficiency rather than low overall deployment cost. The deployment limitation is now stated explicitly in both the experiments and the conclusion.

Under the same inference setting, D3-CalibCount requires 1.484 TFLOPs versus 0.284 TFLOPs for STEERER-HRNet, together with higher latency and memory usage.

Revised text in manuscript: Table 4 reports deployment-oriented metrics under the same inference setting. Compared with STEERER-HRNet, D3-CalibCount retains a smaller trainable part but incurs a larger total model size and higher inference cost because the frozen DINOv3-L backbone dominates the forward pass. Under this setting, D3-CalibCount requires 1.484 TFLOPs, compared with 0.284 TFLOPs for STEERER-HRNet. Combined with the measured latency, FPS, and peak memory, this shows a clear accuracy-cost trade-off in deployment. We accordingly revise our claim from general 'efficiency' to trainable-parameter efficiency and discuss the resulting deployment limitation more explicitly.

The revised experiments also show a clear accuracy-cost trade-off: frozen foundation-feature adaptation improves counting accuracy and preserves training stability, but its current DINOv3-L instantiation is better suited to camera-based intelligent surveillance sensing at GPU-backed monitoring nodes than to strict edge-device deployment.

Comment 7. Clarity and presentation: Some sections (especially SHA and PWSP explanations) remain dense and could benefit from clearer intuitive explanations.

Response. We revised these sections with a more direct explanatory style. The SHA section now begins with an intuitive description of what each module decides, the dual-branch explanation is more concrete, and the PWSP section now states its role as a training-only cue in simpler terms. These additions are meant to reduce density without changing the method itself.

Author Response File: Author Response.pdf