Pith. sign in

REVIEW 5 major objections 5 minor 36 references

Depth-Guided Video Object Counting in Crowded Scenes

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Depth cues, fused through cross-attention and occlusion-aware temporal voting, cut mean absolute counting error by 62.01% on a new RGB-D video counting benchmark.

desk verdict Depth-guided counting with a useful new RGB-D dataset, but the headline MAE reduction is unverifiable until the video-level count annotation is specified. read the letter →

arxiv 2608.06236 v1 pith:2WYZUGPS submitted 2026-08-06 cs.CV cs.AI

Yuanjing Xu , Xinyan Liu , Weidong Chen , Zixuan Zou , Linhao Zhang , Zhuangzhe Meng , Antoni B. Chan , Weigang Zhang This is my paper
classification cs.CVcs.AI
keywords videoobjectcountingRGB-Ddepth-guideddetectionocclusionpredictioncross-framede-duplicationtrackingopen-vocabularycrowdedscenes
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Video object counting in dense scenes fails when similar objects overlap: an RGB-only detector sees one blob where several instances sit, and cross-frame matching fragments or duplicates tracks. This paper tries to establish that synchronized depth breaks that bottleneck, and proposes a depth-guided detector plus a tracking and de-duplication stage that together report a 62.01% reduction in mean absolute counting error over existing baselines on a new RGB-D video counting dataset. The method fuses RGB and depth through cross-attention with a depth-affinity bias, predicts an occlusion score per instance, and uses depth in trajectory matching and temporal voting. If the claim holds, reliable video counting in shelf and warehouse scenes becomes feasible with the depth sensors already common on modern devices.

What carries the argument

The load-bearing object is the Depth Affinity Bias inside a multi-scale RGB-D cross-attention module: RGB features act as queries and depth features as keys and values, and each attention weight is suppressed by the normalized depth difference between the two image positions (Eqs. 1-2), so feature mixing across depth discontinuities is penalized. Two supporting mechanisms carry the argument: a continuous occlusion score per instance, built from spatial overlap and depth ordering (Eq. 17) and supervised through an occlusion prediction head, and a trajectory-matching cost formed from the negative log of a probabilistic match posterior over position, depth residual, appearance cosine similarity, and an IoU gate (Eqs. 6-8). A final temporal support score accepts a trajectory only when detections are both sufficiently numerous and temporally dense within a sliding window (Eq. 10), letting the pipeline recover heavily occluded objects that would otherwise be discarded.

What would settle it

Run the released code and dataset with cross-validated hyperparameters across multiple random seeds; if the 62.01% MAE gap over the RGB baseline does not reproduce, or if the method matches RGB-only error on glass-door beverage scenes, its own reported failure case, then the central claim is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that depth information, when fused at the feature-representation level rather than appended as an extra channel, resolves the failure mode of RGB-only video counting in crowded scenes. The authors build a depth-guided detector in which RGB features query depth features at every scale, with attention penalized by normalized depth difference, and an explicit occlusion head predicts a continuous per-instance occlusion score. A downstream tracking stage incorporates depth residuals into a probabilistic matching cost and applies occlusion-adaptive temporal voting that keeps heavily occluded trajectories alive when their detections are locally dense. The paper reports that this pipeline lowers mean absolute error by 62.01% against existing baselines on the released RGBD-VideoCount dataset, and that the improvement is not simply from adding a depth channel, since depth-aware tracking baselines without the same fusion mechanism perform substantially worse.

Load-bearing premise

The central result assumes the new RGB-D video test set fairly represents the crowded and occluded scenes the method targets, and that the matching and voting hyperparameters were not tuned directly on that test set.

Editorial extensions

If this is right

  • If the reported gains hold, video counting in dense retail and warehouse scenes can rely on depth-ordered instance separation instead of appearance matching alone, directly reducing missed and duplicate counts.
  • Occlusion-adaptive temporal voting means heavily occluded objects can be recovered as long as their detections are locally dense, which addresses the undercounting that RGB-only trackers suffer when objects disappear and reappear.
  • The released RGB-D benchmark, with multiple coexisting categories, per-sequence exemplars, and video-level regions of interest, gives the community a common testbed for depth-assisted counting and de-duplication.
  • The robustness results imply that structured depth degradations such as blur and holes are tolerable in deployment, while random noise remains the main sensor-related risk.
  • The pseudo-depth results imply that RGB-only videos can inherit part of the depth advantage only when the estimated depth preserves sharp instance-level boundaries.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Extension: the same depth-affinity attention could transfer to other dense-instance tasks, such as RGB-D instance segmentation or 3D scene parsing, wherever occlusion ordering is the main source of confusion.
  • Extension: because the paper attributes the pseudo-depth failure to smoothed object boundaries, sharpening predicted depth with RGB edge alignment should recover part of the generalization gap; this is a directly testable consequence of the paper's own diagnosis.
  • Extension: the observed sensitivity to Gaussian and salt-and-pepper depth noise suggests that an explicit depth-confidence weighting in the fusion could extend the method to noisier commodity sensors, a robustness gain the paper does not demonstrate.
  • Extension: the multi-category coexisting scenes in the new dataset make it possible to ask whether depth helps purely geometric separation or also semantic discrimination, by comparing per-category errors on visually confusable pairs such as beverage versus box.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes a depth-guided pipeline for video object counting in crowded scenes, consisting of a Depth-Guided Detector (DG-Det) that fuses RGB and depth features through multi-scale cross-attention with a Depth Affinity Bias and an occlusion prediction head, and a Depth-Guided Tracking and De-duplication module (DG-Track) that performs depth-aware online association and occlusion-adaptive temporal voting. The authors also introduce and release RGBD-VideoCount, a synchronized RGB-D video counting dataset with six categories. The central claim is a 62.01% reduction in MAE relative to existing baselines, with reported test MAE of 7.97 on RGBD-VideoCount in the text+exemplar setting. The paper includes ablations showing contributions of cross-attention fusion, depth affinity bias, occlusion head, depth coverage, and robustness to depth corruptions.

Significance. If the claims hold, the work addresses a genuinely understudied task—video-level individual counting in dense, occluded, multi-category scenes—and provides a useful resource in the form of a new RGB-D dataset with synchronized depth, per-frame boxes, exemplars, and ROIs. The release of source code and dataset is a concrete strength, and the central idea of using depth cues both for single-frame disambiguation and cross-frame de-duplication is well motivated. The main results, however, rest on a dataset whose video-level ground-truth definition is not fully specified, and on comparisons that are not yet sufficiently controlled to support the headline 62.01% reduction. The technical components are described in considerable detail, and the appendix provides derivations for the association cost and the occlusion score, which helps reproducibility beyond the main text.

major comments (5)
  1. [§5.1, §4, Table 1] The video-level ground-truth count y_i used in Sec. 5.1 is not defined by any released annotation. Table 1 lists boxes (B), exemplar boxes (EB), and ROI, but does not list cross-frame instance/track IDs or a per-video unique-object count. Without such identifiers, one cannot compute a unique-object reference count against which the de-duplication and temporal-voting components are evaluated, so the reported 7.97-vs-20.98 MAE comparison in Table 2 is not verifiable. Please either release per-video counts, release cross-frame instance IDs for all annotated boxes, or explicitly state how y_i was derived from the current annotations.
  2. [Table 2, 'Ours' row vs CountVID] The headline 62.01% MAE reduction is computed against the weakest baseline: (20.98−7.97)/20.98 for CountVID in the text+exemplar setting. Against the strongest RGB-D baseline, DepTR-MOT, the reduction is about 44% (14.33 to 7.97), and against RGB-only SparseTrack it is about 65% but still far smaller than the headline figure when baselines are properly matched. More importantly, CountVID uses a different detection front-end (COUNTGD-BOX with SAM 2.1), so the comparison conflates detector improvements with tracker improvements. Please provide controlled comparisons where the same detector (e.g., CountGD-Box) is paired with each tracker, and report the same metrics for all methods with error bars or at least multiple seeds for the main table.
  3. [Eqs. (6)–(10), §3.2] The association cost and temporal voting depend on several hyperparameters whose values are not disclosed: sigma_p, sigma_d, kappa_f, lambda_iou, tau_iou, H_base, N_min, rho, and W. The sensitivity analysis only varies beta (Table 5) and compares fixed-threshold voting with adaptive voting; it does not vary the post-IoU gate, the matching-cost weights, or the lifespan and density thresholds. Since the de-duplication stage is a central contribution, please provide the chosen values for all these parameters and a sensitivity study over the ones that most affect the number of valid trajectories, for example tau_iou and N_min.
  4. [Tables 3 and 4] The two ablation tables report inconsistent RGB-only baselines. Table 3 lists 'RGB-only (0%)' with validation MAE/RMSE 19.88/28.46 and test MAE/RMSE 15.47/25.26, whereas Table 4 lists 'RGB-only' with validation 18.72/32.40 and test 18.07/26.80. If both are meant to be the same configuration (same detector, same tracking pipeline), the discrepancies need to be explained; if the configurations differ, the captions must say how. This inconsistency makes it difficult to assess the incremental gain of each component.
  5. [§E, Table 7] The generalization evaluation on FSCD-147 with pseudo-depth shows that the proposed method is worse than CountGD-Box in the 'both' prompt setting (test MAE 11.91 vs. 10.29, and lower AP/AP50). The text explains that pseudo-depth has limited quality, but the abstract and introduction frame the method as a general depth-guided counting pipeline. Please state this limitation explicitly in the main text and temper the generalization claim, or provide evidence that the method improves over CountGD-Box when pseudo-depth is supplied as input in a comparable setting.
minor comments (5)
  1. [Abstract, §4] The abstract contains 'persequence' instead of 'per sequence', and Table 1 reports '195+' videos while the surrounding text says 195 video clips; please make the counts consistent.
  2. [Eq. (4)] In Eq. (4), the scaling parameter gamma is generated from global average pooling while the shift parameter beta is generated from global max pooling; if this asymmetry is intentional, please state why, since using different pooling for the two FiLM parameters is not self-evident.
  3. [Table 7 caption] The caption for Table 7 says 'best and second-best results are marked in the unifiedbothsetting' but the table does not visually mark any results; please clarify the intended formatting or remove the sentence.
  4. [§5.3, Table 3] The partial-depth settings (25% and 75%) show RMSE values much larger than the RGB-only baseline (e.g., test RMSE 35.85 for 25% depth vs. 25.26 for RGB-only). A short explanation of why partial depth can hurt RMSE while improving MAE would help, since this is not obvious from the text.
  5. [§2, Related Work] The related-work section cites DepthMOT and DepTR-MOT as RGB-D trackers, but the experimental comparison with them in Table 2 uses their reported numbers without error bars or a description of input resolution and annotation format; please state whether these models were retrained or evaluated with the released checkpoints and the same ROI filtering.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported MAE reduction is an empirical benchmark comparison, and no prediction reduces to a fitted constant or self-citation by construction.

full rationale

The paper's central claim, a 62.01% reduction in MAE on RGBD-VideoCount (20.98 to 7.97 in Table 2), is an empirical result computed from predicted counts against ground-truth video counts; it is not derived from an equation whose inputs already contain the claimed output. The occlusion-aware head is supervised by a continuous occlusion score constructed from boxes and depth (Appendix B), but that score is a training label, not the counting target y_i, and the predicted occlusion only modulates trajectory voting (Eq. 10). The association cost in Eq. 8 is derived from explicit probabilistic likelihoods with stated independence assumptions; even though some hyperparameters are not fully reported, that is a reproducibility issue, not evidence that the cost is fitted to the test counts. Self-citations (e.g., [19, 22-24]) appear only in related-work context and are not load-bearing for the proposed method or its evaluation. The self-collected dataset and the absence of explicitly stated cross-frame instance IDs in Table 1 are legitimate verification and reproducibility concerns, but they do not make the derivation circular: the ground-truth count is not defined in terms of the model's trajectory voting. The inclusion of an external benchmark (FSCD-147) where the method does not claim superiority further confirms that the main result is not forced by construction. The probabilistic derivation in Appendix C is a formal rewriting of the matching cost, not a source from which the counting outcome is extracted as an input.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a set of undisclosed hyperparameters in the tracking and voting stages, plus assumptions about depth quality and alignment. The paper does not introduce new physical entities or theoretical constructs.

free parameters (6)
  • Depth Affinity Bias coefficient β_bias = 1.0 (default, range 0.5-3.0 tested)
    Controls the strength of depth-based suppression in attention; chosen by hand with sensitivity analysis in Table 5.
  • Tracking noise scales σ_p, σ_d and appearance concentration κ_f = not disclosed
    Define the association cost weights in Eq. 9; values are not given and likely tuned on validation.
  • IoU gate threshold τ_iou and penalty λ_iou = not disclosed
    Used in the geometric prior in Eq. 8; not specified in the paper.
  • Voting parameters H_base, N_min, ρ, W = not disclosed
    Control trajectory validity in Eq. 10; values are not provided.
  • Loss weights λ1-λ4 = not disclosed
    Used in Eq. 13 for training; not specified.
  • Occlusion score parameters δ_z, τ_z = not disclosed
    Used in Eq. 15 to build occlusion supervision; not specified.
assumptions (4)
  • domain assumption Depth ordering implies occlusion ordering
    The occlusion score and depth affinity bias assume that an object closer to the camera occludes a farther one; this fails for transparent or reflective surfaces, as acknowledged in Sec. F.
  • standard math Conditional independence of position, depth, and appearance in matching
    Used to write the association posterior as a product in Eq. 7; a modeling convenience without empirical validation.
  • domain assumption Pixel-aligned RGB and depth streams
    The method assumes synchronized and aligned RGB-D input, as stated in Sec. 3; any misalignment would break the cross-attention fusion.
  • domain assumption The ROI is correctly specified and covers all counting targets
    Counting is limited to a given ROI; in practice, ROI errors would bias the count.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Depth-Guided Video Object Counting in Crowded Scenes." pith.science (2026). https://pith.science/paper/2WYZUGPS

@misc{pith2026260806236,
  author       = {Pith},
  title        = {Pith review of: Depth-Guided Video Object Counting in Crowded Scenes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2WYZUGPS}},
  note         = {Machine review of arXiv:2608.06236}
}
read the original abstract

Our primary objective is to advance video object counting in crowded scenes, aiming to robustly count all instances of a target category based on given text or visual prompts. Existing methods rely on RGB information, limiting their discriminative ability in crowded and occluded conditions. To address this, we propose a Depth-Guided Detector (DG-Det) along with a general post-processing pipeline. By integrating depth cues with multi-scale RGB-D cross-attention and explicit occlusion prediction, our method enhances spatial understanding and achieves robust detection in crowded and occluded scenes. Furthermore, we introduce a unified de-duplication framework to eliminate cross-frame redundant counting. To facilitate future research, we also release a new RGB-D Video Object Counting dataset featuring depth information and multiple object categories persequence. Extensive experiments demonstrate that our method achieves a 62.01\% reduction in MAE compared to existing baselines, and also produces consistent improvements in RMSE. We provide the source code at https://github.com/streamer-AP/DG-Net and the dataset at https://huggingface.co/datasets/aerospace123/RGBD-VideoCount.

Figures

Figures reproduced from arXiv: 2608.06236 by the authors.

Figure 1
Figure 1. Qualitative comparison in a dense shelf scene. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of our framework. Top: the Depth-Guided Detector (DG-Det) takes a text and visual query and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison between predicted pseudo [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Example predictions of our method in two typical challenging scenarios. For each video, several sampled frames and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: (a) Beverage scene: Demonstrates the model’s performance in occluded scenarios. (b) Cabinet scene: Demonstrates the [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 7
Figure 7. Figure 7: Failure cases in glass-door beverage cabinet scenes. The first and third rows show the RGB images overlaid with [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 6
Figure 6. Figure 6: Qualitative examples on FSCD-147 with pseudo-depth. Top row: input images with predicted counts (Pred) and [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 24 canonical work pages

  1. [1]

    Niki Amini-Naieni, Kiana Amini-Naieni, Tengda Han, and Andrew Zisserman

    work page

  2. [2]

    Niki Amini-Naieni, Tengda Han, and Andrew Zisserman. 2024. Countgd: Multi- modal open-world counting.Advances in Neural Information Processing Systems 37 (2024), 48810–48837

    work page 2024

  3. [3]

    Niki Amini-Naieni and Andrew Zisserman. 2026. Open-world object counting in videos. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 2300–2308

    work page 2026

  4. [4]

    Carlos Arteta, Victor Lempitsky, J Alison Noble, and Andrew Zisserman. 2014. Interactive object counting. InEuropean conference on computer vision. Springer, 504–518

    work page 2014

  5. [5]

    Olga Barinova, Victor Lempitsky, and Pushmeet Kohli. 2010. On detection of multiple object instances using hough transforms. In2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. IEEE, 2233–2240

    work page 2010

  6. [6]

    Siu-Yeung Cho, Tommy WS Chow, and Chi-Tat Leung. 1999. A neural-based crowd estimation by hybrid global learning algorithm.IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics)29, 4 (1999), 535–541. 8

    work page 1999

  7. [7]

    Siyang Dai, Jun Liu, and Ngai-Man Cheung. 2024. Referring expression count- ing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 16985–16995

    work page 2024

  8. [8]

    Buyin Deng, Lingxin Huang, Kai Luo, Fei Teng, and Kailun Yang. 2025. DepTR- MOT: Unveiling the Potential of Depth-Informed Trajectory Refinement for Multi-Object Tracking.arXiv preprint arXiv:2509.17323(2025)

    arXiv 2025

Show all 36 references
  1. [9]

    Chaitanya Desai, Deva Ramanan, and Charless C Fowlkes. 2011. Discriminative models for multi-class object layout.International journal of computer vision95, 1 (2011), 1–12

    2011

  2. [10]

    Guangshuai Gao, Junyu Gao, Qingjie Liu, Qi Wang, and Yunhong Wang. 2020. Cnn-based density estimation and crowd counting: A survey.arXiv preprint arXiv:2003.12783(2020)

    2020 arXiv

  3. [11]

    Shenjian Gong, Shanshan Zhang, Jian Yang, Dengxin Dai, and Bernt Schiele

  4. [12]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition. 770–778

    2016

  5. [13]

    Meng-Ru Hsieh, Yen-Liang Lin, and Winston H Hsu. 2017. Drone-based object counting by spatially regularized regional proposal network. InProceedings of the IEEE international conference on computer vision. 4145–4153

    2017

  6. [14]

    Ruixiang Jiang, Lingbo Liu, and Changwen Chen. 2023. Clip-count: Towards text- guided zero-shot object counting. InProceedings of the 31st ACM International Conference on Multimedia. 4535–4545

    2023

  7. [15]

    Dan Kong, Douglas Gray, and Hai Tao. 2006. A viewpoint invariant approach for crowd counting. In18th International Conference on Pattern Recognition (ICPR’06), Vol. 3. IEEE, 1187–1190

    2006

  8. [16]

    Bo Li, Hongbo Huang, Ang Zhang, Peiwen Liu, and Cheng Liu. 2021. Approaches on crowd counting and density estimation: a review.Pattern Analysis and Appli- cations24, 3 (2021), 853–874

    2021

  9. [17]

    Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. 2017. Feature pyramid networks for object detection. InProceed- ings of the IEEE conference on computer vision and pattern recognition. 2117–2125

    2017

  10. [18]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. InEuropean conference on computer vision. Springer, 740–755

    2014

  11. [19]

    Wei Lin and Antoni B Chan. 2024. A fixed-point approach to unified prompt- based counting. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 3468–3476

    2024

  12. [20]

    Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. 2022. Countr: Transformer-based generalised visual counting.arXiv preprint arXiv:2208.13721 (2022)

    2022 arXiv

  13. [21]

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. 2024. Grounding dino: Marry- ing dino with grounded pre-training for open-set object detection. InEuropean conference on computer vision. Springer, 38–55

    2024

  14. [22]

    Xinyan Liu, Guorong Li, Zhenjun Han, Weigang Zhang, Yifan Yang, Qingming Huang, and Nicu Sebe. 2021. Exploiting sample correlation for crowd count- ing with multi-expert network. In2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 3195–3204

    2021

  15. [23]

    Xinyan Liu, Guorong Li, Yuankai Qi, Zhenjun Han, Anton van den Hengel, Nicu Sebe, Ming-Hsuan Yang, and Qingming Huang. 2024. Consistency-aware anchor pyramid network for crowd localization.IEEE transactions on pattern analysis and machine intelligence(2024)

    2024

  16. [24]

    Xinyan Liu, Guorong Li, Yuankai Qi, Ziheng Yan, Zhenjun Han, Anton Van Den Hengel, Ming-Hsuan Yang, and Qingming Huang. 2024. Weakly supervised video individual counting. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 19228–19237

    2024

  17. [25]

    Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. 2022. Few-shot object counting and detection. InEuropean Conference on Computer Vision. Springer, 348–365

    2022

  18. [26]

    Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. 2018. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, Vol. 32

    2018

  19. [27]

    Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. 2021. Learning to count everything. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 3394–3403

    2021

  20. [28]

    Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. 2021. Rethinking counting and localization in crowds: A purely point-based framework. InProceedings of the IEEE/CVF international conference on computer vision...

    2021

  21. [29]

    Nikola Ðukić, Alan Lukežič, Vitjan Zavrtanik, and Matej Kristan. 2023. A low- shot object counting network with iterative prototype adaptation. InProceedings of the IEEE/CVF International Conference on Computer Vision. 18872–18881

    2023

  22. [30]

    Jiapeng Wu and Yichen Liu. 2024. Depthmot: Depth cues lead to a strong multi- object tracker.arXiv preprint arXiv:2404.05518(2024)

    2024 arXiv

  23. [31]

    Shaokai Wu and Fengyu Yang. 2023. Boosting detection in crowd analysis via underutilized output features. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 15609–15618

    2023

  24. [32]

    Kelei Xu, Chunyan Wang, Wanzhong Zhao, and Jinqiang Liu. 2025. DGT: Depth- guided RGB-D occluded target detection with transformers.Applied Intelligence 55, 5 (2025), 300

    2025

  25. [33]

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. 2024. Depth anything v2.Advances in Neural Information Processing Systems37 (2024), 21875–21911

    2024

  26. [34]

    Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. 2023. MonoDETR: Depth-Guided Transformer for Monocular 3D Object Detection. In2023 IEEE/CVF International Conference on Computer Vision (ICCV). 9121–9132. doi:10.1109/ICCV51070.2023.008...

    2023

  27. [2022]

    InEuropean Conference on Computer Vision

    Class-agnostic object counting robust to intraclass diversity. InEuropean Conference on Computer Vision. Springer, 388–403

  28. [2023]

    Open-world text-specified object counting.arXiv preprint arXiv:2306.01851 (2023)

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.