Pith. sign in

REVIEW 2 major objections 5 minor 43 references

MSMVD: Exploiting Multi-scale Image Features via Multi-scale BEV Features for Multi-view Pedestrian Detection

T0 review · 2 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read MSMVD shows that carrying multi-scale image features through the BEV projection, rather than a single image scale, is what lets end-to-end multi-view pedestrian detectors handle small, large, and mixed-scale pedestrians, raising MODA by 4.5

desk verdict Solid architectural contribution with convincing ablations; the exact 4.5-point SOTA margin is unverified because the baseline re-implementations are unreleased and M-MVOT is missing from the GMVD comparison. read the letter →

arxiv 2508.20447 v1 pith:WSRG3MQB submitted 2025-08-28 cs.CV

Taiga Yamane , Satoshi Suzuki , Ryo Masumura , Shota Orihashi , Tomohiro Tanaka , Mana Ihori , Naoki Makishima , Naotaka Kawata This is my paper · ORCID
classification cs.CV
keywords multi-viewpedestriandetectionbird'seyeviewmulti-scaleimagefeaturesBEVfeaturepyramidnetworkoccupancymapGMVDscalevariation
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

MSMVD is built on a diagnosis: end-to-end multi-view pedestrian detectors that form a single bird's-eye-view (BEV) feature from one image scale systematically miss pedestrians who appear small or large in every view, or whose apparent size differs sharply between views. The paper proposes fixing this by projecting the encoder's multi-scale image features into BEV space scale-by-scale, generating multi-scale BEV features, and then fusing those features with a BEV feature pyramid network before predicting occupancy maps at each scale. On the GMVD benchmark, this raises multiple object detection accuracy (MODA) from the previous best 75.7 to 80.2, a 4.5-point gain, with smaller gains on Wildtrack and MultiviewX. If the result holds, scale handling in multi-view detection is not just an image-side detail: the scale information must be carried through the BEV projection itself.

What carries the argument

The central mechanism is MSP (multi-scale projection), which maps each scale's enhanced image features into a BEV grid whose resolution stays proportional to the image scale, using camera parameters and projections at five heights so scale-specific information survives the view-to-BEV transform. The second component is BEV-FPN, a path aggregation network over the BEV scales that fuses information across those scales; the fused multi-scale BEV features then feed per-scale occupancy maps merged at inference.

What would settle it

Run the official released checkpoints of the five re-implemented methods on GMVD under the same evaluation protocol; if any scores within 4.5 MODA of MSMVD, the headline margin is not against the true previous best. Separately, fix all image features to a single scale inside MSMVD; if MODA stays near 80.2, the multi-scale mechanism is not what carries the gain.

Watch

Extended reading notes

Core claim

Previous end-to-end MVPD methods produce a single-scale BEV feature by projecting one layer of image features; MSMVD instead projects image features from three ResNet stages, after an image-FPN, into BEV space scale-by-scale, max-pools across views at each scale, and runs a PAFPN-style top-down/bottom-up BEV-FPN to combine scale information from multiple views. Multi-scale occupancy maps are predicted and averaged at inference. The paper reports that the multi-scale BEV representation accounts for most of the gain: adding MSP raises MODA on GMVD from 71.8 to 74.9, and adding BEV-FPN raises it to 80.2; MSMVD beats the previous best by 4.5 MODA on GMVD, 0.5 on Wildtrack, and 0.5 on MultiviewX.

Load-bearing premise

The claimed 4.5-point lead over previous state of the art assumes the authors' re-implementations of SHOT, 3DROM, MV Aug, OmniOcc, and MVFP on GMVD are at least as strong as the official released versions; if any is undertuned, the gap is inflated, although the internal baseline-to-MSMVD ablation (71.8 to 80.2) is independent of that comparison.

Editorial extensions

If this is right

  • Multi-scale image features become useful for MVPD only when carried into BEV space at matched resolutions; projecting a single scale forfeits the gain.
  • Detectors using MSMVD's multi-scale BEV features can detect pedestrians missed by single-scale BEV methods in both consistent-scale and cross-view-scale situations.
  • The 4.5-point MODA gain on GMVD is not simply a parameter-count effect: MSMVD with ResNet18 beats the baseline with ResNet101 (80.2 vs. 75.8 MODA).
  • The method transfers across datasets, setting a new state of the art on Wildtrack and MultiviewX as well as GMVD.

Reading between the lines

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

  • The scale-proportional BEV resolution choice is likely a general design lesson: any multi-view detector fusing image features of different resolutions should preserve their resolution ratio in the BEV target rather than upsample everything to one grid.
  • A testable extension would evaluate MSMVD on real-world multi-camera setups with more extreme camera-height or field-of-view differences; the modest Wildtrack gain suggests the benefit may grow with the degree of scale diversity in the data.
  • The multi-scale occupancy merging rule could be replaced by learned fusion or per-scale confidence weighting; since even single-scale inference from the multi-scale BEV features already beats prior methods, the BEV representation, not the averaging rule, likely carries most of the gain.
  • The auxiliary losses on the unused O4 and O5 offset maps contribute 0.6 MODA, hinting that multi-scale auxiliary supervision could be a cheap addition to other BEV detection heads.
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

2 major / 5 minor

Summary. The paper proposes MSMVD, a multi-view pedestrian detection method that generates multi-scale BEV features by projecting multi-scale image features (from an image-FPN) into the BEV space scale-by-scale, aggregating them across views via max pooling, and fusing them with a BEV-FPN. Multi-scale occupancy maps are predicted and merged at inference. On GMVD the method reports a 4.5-point MODA improvement over the previous best (80.2 vs 75.7), and smaller gains on Wildtrack and MultiviewX. Ablations show monotone gains from adding MSP and BEV-FPN, and backbone-size controls indicate the gains are not simply parameter-count effects.

Significance. If the reported results hold, MSMVD provides a simple and effective architectural response to a known weakness of end-to-end MVPD methods: the inability to handle pedestrians with consistently small/large scales or vastly different scales across views. The paper's strengths are its clear architecture, the clean internal ablation chain (Table 3a: 71.8 to 74.9 to 80.2), the backbone-size control (Table 4), and the supplementary ablations covering offset prediction, pooling, bottom-up paths, and the scaling factor. These give credible evidence that the proposed components themselves, not just additional parameters, drive the improvement. However, the headline GMVD state-of-the-art claim rests on a baseline table in which five of seven prior methods are author re-implementations, and a strong existing method (M-MVOT) is omitted from the GMVD comparison. The internal contribution is defensible, but the external SOTA margin is not yet fully verified.

major comments (2)
  1. [§4.3, Table 1] The GMVD SOTA claim depends on author re-implementations of SHOT, 3DROM, MV Aug, OmniOcc, and MVFP, but the paper provides no code, hyperparameters, or training details for these re-implementations. More importantly, M-MVOT, which appears in Table 2 with strong Wildtrack/MultiviewX results, is not included in the GMVD table. The text says methods evaluated only on Wildtrack/MultiviewX were re-implemented for GMVD; M-MVOT satisfies that criterion. Without adding M-MVOT (and ideally validating the re-implementations against official results on at least one dataset), the 4.5-point margin over the previous highest MODA is not established. The internal ablation remains valid, but the external comparison needs strengthening.
  2. [§4.2, Tables 1-4] All reported metrics are single-run point estimates with no error bars or multiple-seed statistics. For a 4.5-point claim over a baseline set that includes unreleased re-implementations, some indication of run-to-run variance (or at least a statement that fluctuations are negligible) is needed to distinguish signal from noise.
minor comments (5)
  1. [Supplementary, Table 7] Typo in the table header: BEV-FPV should be BEV-FPN.
  2. [§4.1 and Table 1] The name Vora is rendered as 'V ora' in several places (e.g., 'V ora+' in Table 1 and the text). Please fix the spacing.
  3. [§4.3, Table 2] The dagger for OmniOcc in Table 2 is explained only in the text; the caption should state explicitly that OmniOcc is re-implemented by the authors, matching the convention in Table 1.
  4. [§3.4, Eq. (3)] The inference merging uses equal weights 1/3 for the three scales. The authors do not ablate alternative weights (e.g., learned or scale-dependent weights). This is a minor omission, but it would strengthen the argument that simple averaging is the right choice.
  5. [§4.2] The detection threshold is fixed at 0.4 without a sensitivity analysis. Since final metrics depend on this threshold, a short sentence or supplementary figure showing robustness to threshold choice would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the architecture is evaluated against fixed external benchmarks, and the only self-citation is a non-load-bearing baseline.

full rationale

MSMVD's contribution is an architectural change (multi-scale projection plus BEV-FPN) whose headline gain is measured on held-out test splits of GMVD, Wildtrack, and MultiviewX. No parameter is fitted to the test sets, and the per-component ablations (Table 3a: baseline 71.8 → +MSP 74.9 → +BEV-FPN 80.2) show the improvement is produced by the proposed modules rather than by a re-fitted constant. The only self-citation, Suzuki+ [33], is an earlier method by overlapping authors used purely as a comparison baseline in Table 1; its removal would not change any equation or architectural derivation. The daggered re-implementations (SHOT†, 3DROM†, MV Aug†, OmniOcc†, MVFP†) are a legitimate threat to the strength of the 'previous highest' comparison, but that is a baseline-completeness/tuning risk, not circularity: no result is defined in terms of itself and no prior theorem is imported from the authors' own work. The use of FPN/PAFPN is taken from external monocular detection literature and applied directly, not smuggled in via self-citation. No Eq. (1)-(3) reduces to its input by construction, and no 'prediction' is a renamed fitted parameter. Thus the derivation chain is self-contained, with a score of 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities. Its central claim rests on hand-chosen architectural hyperparameters and standard domain assumptions about camera calibration, pedestrian scale, and BEV projection. The most performance-sensitive free choices are the per-scale Gaussian kernel diameters, the MSP resolution mapping, and the merge weights.

free parameters (5)
  • Gaussian kernel diameters for ground-truth occupancy maps = 20, 10, 5 pixels for M3, M4, M5
    Hand-set in Supp. A. These diameters define the soft ground-truth targets at each scale and directly affect the loss and the final detections.
  • Detection threshold = 0.4
    Used at inference to convert occupancy scores into pedestrian detections; affects MODA and recall.
  • Multi-scale occupancy merge weights = 1/3 each
    Equation (3) averages M3, M4, and M5 with equal weight. Alternative weighting could change the final detection map.
  • MSP scale factor gamma_n = Resolutions X/2^(l-2) by Y/2^(l-2)
    Chosen so BEV resolution ratios match image feature resolution ratios. Supp. B shows the alternative of equal resolutions lowers MODA by 1.8 points, so this choice is performance-relevant.
  • Projection height planes = z_i = 30 * i cm for i = 0..4
    Five height planes inherited from 3DROM [30]. The number and spacing of planes are hand-chosen and affect what information is projected into the BEV feature.
assumptions (5)
  • domain assumption Camera intrinsics and extrinsics (K, R, T) are known and calibrated for every view.
    Equation (1) projects image pixels to world coordinates using K[R|T]; the entire BEV projection pipeline depends on this.
  • domain assumption Pedestrians can be represented by projecting image features onto discrete horizontal planes at heights z0..z4.
    MSP projects each image feature at five heights and concatenates results. This assumes the height discretization captures the pedestrian body sufficiently.
  • domain assumption High-resolution image features better represent small pedestrians and low-resolution features better represent large pedestrians.
    The motivation in Section 1 and 3.1 relies on this standard multi-scale detection assumption; the paper does not verify it independently in BEV space.
  • domain assumption Max pooling across views preserves the most informative view features at each scale.
    Aggregation of projected features {P_l^n} uses max pooling along the view direction. Supp. B shows max pooling outperforms mean pooling on GMVD.
  • domain assumption PAFPN-style top-down and bottom-up fusion is an effective way to combine multi-scale BEV features.
    BEV-FPN is directly adapted from PAFPN; Supp. B shows removing the bottom-up path lowers MODA, but the overall effectiveness of the fusion design is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MSMVD: Exploiting Multi-scale Image Features via Multi-scale BEV Features for Multi-view Pedestrian Detection." pith.science (2026). https://pith.science/paper/WSRG3MQB

@misc{pith2026250820447,
  author       = {Pith},
  title        = {Pith review of: MSMVD: Exploiting Multi-scale Image Features via Multi-scale BEV Features for Multi-view Pedestrian Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WSRG3MQB}},
  note         = {Machine review of arXiv:2508.20447}
}
abstract

Multi-View Pedestrian Detection (MVPD) aims to detect pedestrians in the form of a bird's eye view (BEV) from multi-view images. In MVPD, end-to-end trainable deep learning methods have progressed greatly. However, they often struggle to detect pedestrians with consistently small or large scales in views or with vastly different scales between views. This is because they do not exploit multi-scale image features to generate the BEV feature and detect pedestrians. To overcome this problem, we propose a novel MVPD method, called Multi-Scale Multi-View Detection (MSMVD). MSMVD generates multi-scale BEV features by projecting multi-scale image features extracted from individual views into the BEV space, scale-by-scale. Each of these BEV features inherits the properties of its corresponding scale image features from multiple views. Therefore, these BEV features help the precise detection of pedestrians with consistently small or large scales in views. Then, MSMVD combines information at different scales of multiple views by processing the multi-scale BEV features using a feature pyramid network. This improves the detection of pedestrians with vastly different scales between views. Extensive experiments demonstrate that exploiting multi-scale image features via multi-scale BEV features greatly improves the detection performance, and MSMVD outperforms the previous highest MODA by $4.5$ points on the GMVD dataset.

Figures

Figures reproduced from arXiv: 2508.20447 by the authors.

Figure 1
Figure 1. (a) Previous end-to-end MVPD methods generate a single-scale BEV feature from [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (Top) Comparison of the predicted BEV maps using MVFP [ [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 36 canonical work pages

  1. [1]

    Enhancing multi-view pedestrian detection through generalized 3D feature pulling

    Sithu Aung, Haesol Park, Hyungjoo Jung, and Junghyun Cho. Enhancing multi-view pedestrian detection through generalized 3D feature pulling. In WACV, 2024

    work page 2024

  2. [2]

    Multi-view pedestrian occupancy prediction with a novel synthetic dataset

    Sithu Aung, Min-cheol Sagong, and Junghyun Cho. Multi-view pedestrian occupancy prediction with a novel synthetic dataset. In AAAI, 2025

    work page 2025

  3. [3]

    Deep occlusion reasoning for multi- camera multi-target detection

    Pierre Baqué, François Fleuret, and Pascal Fua. Deep occlusion reasoning for multi- camera multi-target detection. In ICCV, 2017. Y AMANE ET AL.: MSMVD 11

    work page 2017

  4. [4]

    Deep multi-camera people detection

    Tatjana Chavdarova and François Fleuret. Deep multi-camera people detection. In ICMLA, 2017

    work page 2017

  5. [5]

    Wildtrack: A multi-camera hd dataset for dense unscripted pedestrian detection

    Tatjana Chavdarova, Pierre Baqué, Stéphane Bouquet, Andrii Maksai, Cijo Jose, Timur Bagautdinov, Louis Lettry, Pascal Fua, Luc Van Gool, and François Fleuret. Wildtrack: A multi-camera hd dataset for dense unscripted pedestrian detection. In CVPR, 2018

    work page 2018

  6. [6]

    YOLO-MS: rethinking multi-scale representation learning for real-time object detection

    Yuming Chen, Xinbin Yuan, Jiabao Wang, Ruiqi Wu, Xiang Li, Qibin Hou, and Ming- Ming Cheng. YOLO-MS: rethinking multi-scale representation learning for real-time object detection. TPAMI, 2025

    work page 2025

  7. [7]

    SportsMOT: A large multi-object tracking dataset in multiple sports scenes

    Yutao Cui, Chenkai Zeng, Xiaoyu Zhao, Yichun Yang, Gangshan Wu, and Limin Wang. SportsMOT: A large multi-object tracking dataset in multiple sports scenes. In ICCV, 2023

    work page 2023

  8. [8]

    Histograms of oriented gradients for human detection

    Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In CVPR, 2005

    work page 2005

Show all 43 references
  1. [9]

    ImageNet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009

    2009

  2. [10]

    Pedestrian detection: An evaluation of the state of the art

    Piotr Dollar, Christian Wojek, Bernt Schiele, and Pietro Perona. Pedestrian detection: An evaluation of the state of the art. TPAMI, 2011

    2011

  3. [11]

    Multi-object detection and tracking (MODT) machine learning model for real-time video surveillance systems

    Mohamed Elhoseny. Multi-object detection and tracking (MODT) machine learning model for real-time video surveillance systems. CSSP, 2020

    2020

  4. [12]

    Two-level data augmen- tation for calibrated multi-view detection

    Martin Engilberge, Haixin Shi, Zhiye Wang, and Pascal Fua. Two-level data augmen- tation for calibrated multi-view detection. In WACV, 2023

    2023

  5. [13]

    Multicamera people tracking with a probabilistic occupancy map

    Francois Fleuret, Jerome Berclaz, Richard Lengagne, and Pascal Fua. Multicamera people tracking with a probabilistic occupancy map. TPAMI, 2007

    2007

  6. [14]

    YOLOX: Exceeding yolo series in 2021

    Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: Exceeding yolo series in 2021. arXiv:2107.08430, 2021

    2021 arXiv

  7. [15]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016

    2016

  8. [16]

    Multiview detection with shadow transformer (and view-coherent data augmentation)

    Yunzhong Hou and Liang Zheng. Multiview detection with shadow transformer (and view-coherent data augmentation). In ACMMM, 2021

    2021

  9. [17]

    Multiview detection with feature perspective transformation

    Yunzhong Hou, Liang Zheng, and Stephen Gould. Multiview detection with feature perspective transformation. In ECCV, 2020

    2020

  10. [18]

    Booster-SHOT: Boosting stacked ho- mography transformations for multiview pedestrian detection with attention

    Jinwoo Hwang, Philipp Benz, and Pete Kim. Booster-SHOT: Boosting stacked ho- mography transformations for multiview pedestrian detection with attention. InWACV, 2024

    2024

  11. [19]

    Framework for performance evaluation of face, text, and vehicle detection and tracking in video: Data, metrics, and protocol

    Rangachar Kasturi, Dmitry Goldgof, Padmanabhan Soundararajan, Vasant Manohar, John Garofolo, Rachel Bowers, Matthew Boonstra, Valentina Korzhova, and Jing Zhang. Framework for performance evaluation of face, text, and vehicle detection and tracking in video: Data, metrics, and...

    2008

  12. [20]

    F2DNet: Fast focal detection network for pedestrian detection

    Abdul Hannan Khan, Mohsin Munir, Ludger van Elst, and Andreas Dengel. F2DNet: Fast focal detection network for pedestrian detection. In ICPR, 2022

    2022

  13. [21]

    Localized semantic feature mixers for efficient pedestrian detection in autonomous driving

    Abdul Hannan Khan, Mohammed Shariq Nawaz, and Andreas Dengel. Localized semantic feature mixers for efficient pedestrian detection in autonomous driving. In CVPR, 2023

    2023

  14. [22]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017

    2017

  15. [23]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. In ICCV, 2017

    2017

  16. [24]

    Path aggregation network for instance segmentation

    Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, 2018

    2018

  17. [25]

    SSD: Single shot multibox detector

    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng- Yang Fu, and Alexander C Berg. SSD: Single shot multibox detector. In ECCV, 2016

    2016

  18. [26]

    Center and scale prediction: Anchor-free approach for pedestrian and face detection

    Wei Liu, Irtiza Hasan, and Shengcai Liao. Center and scale prediction: Anchor-free approach for pedestrian and face detection. PR, 2023

    2023

  19. [27]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In ICLR, 2018

    2018

  20. [28]

    Distinctive image features from scale-invariant keypoints

    David G Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 2004

    2004

  21. [29]

    RTMDet: An empirical study of designing real-time object detectors

    Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. RTMDet: An empirical study of designing real-time object detectors. arXiv:2212.07784, 2022

    2022 arXiv

  22. [30]

    3D random occlusion and multi-layer projection for deep multi-camera pedestrian localization

    Rui Qiu, Ming Xu, Yuyao Yan, Jeremy S Smith, and Xi Yang. 3D random occlusion and multi-layer projection for deep multi-camera pedestrian localization. In ECCV, 2022

    2022

  23. [31]

    Conditional random fields for multi-camera object detection

    Gemma Roig, Xavier Boix, Horesh Ben Shitrit, and Pascal Fua. Conditional random fields for multi-camera object detection. In ICCV, 2011

    2011

  24. [32]

    Stacked homography transformations for multi-view pedestrian detection

    Liangchen Song, Jialian Wu, Ming Yang, Qian Zhang, Yuan Li, and Junsong Yuan. Stacked homography transformations for multi-view pedestrian detection. In ICCV, 2021

    2021

  25. [33]

    Scene generalized multi-view pedestrian detection with rotation-based augmentation and regularization

    Satoshi Suzuki, Shotaro Tora, and Ryo Masumura. Scene generalized multi-view pedestrian detection with rotation-based augmentation and regularization. In ICIP, 2024

    2024

  26. [34]

    EfficientDet: Scalable and efficient object detection

    Mingxing Tan, Ruoming Pang, and Quoc V Le. EfficientDet: Scalable and efficient object detection. In CVPR, 2020

    2020

  27. [35]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need. In NIPS, 2017

    2017

  28. [36]

    Bringing generalization to deep multi-view pedestrian detection

    Jeet V ora, Swetanjal Dutta, Kanishk Jain, Shyamgopal Karthik, and Vineet Gandhi. Bringing generalization to deep multi-view pedestrian detection. In WACV, 2023. Y AMANE ET AL.: MSMVD 13

    2023

  29. [37]

    A multi modal people tracker for real time human robot interaction

    Tim Wengefeld, Steffen Müller, Benjamin Lewandowski, and Horst-Michael Gross. A multi modal people tracker for real time human robot interaction. In RO-MAN, 2019

    2019

  30. [38]

    Multi-view people tracking via hierarchical trajectory composition

    Yuanlu Xu, Xiaobai Liu, Yang Liu, and Song-Chun Zhu. Multi-view people tracking via hierarchical trajectory composition. In CVPR, 2016

    2016

  31. [39]

    Mahalanobis distance-based multi-view optimal transport for multi-view crowd localization

    Qi Zhang, Kaiyi Zhang, Antoni B Chan, and Hui Huang. Mahalanobis distance-based multi-view optimal transport for multi-view crowd localization. In ECCV, 2024

    2024

  32. [40]

    CityPersons: A diverse dataset for pedestrian detection

    Shanshan Zhang, Rodrigo Benenson, and Bernt Schiele. CityPersons: A diverse dataset for pedestrian detection. In CVPR, 2017

    2017

  33. [41]

    DETRs beat YOLOs on real-time object detection

    Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. DETRs beat YOLOs on real-time object detection. In CVPR, 2024

    2024

  34. [42]

    Objects as points

    Xingyi Zhou, Dequan Wang, and Philipp Krähenbühl. Objects as points. arXiv:1904.07850, 2019

    1904 arXiv

  35. [43]

    Deformable DETR: Deformable transformers for end-to-end object detection

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: Deformable transformers for end-to-end object detection. In ICLR, 2021. Y AMANE ET AL.: MSMVD 1 Supplementary Material MODA MODP Precision Recall Train w/oO4 andO5 79.6 81.0 95.4 83.6 Trai...

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.