Pith. sign in

REVIEW 3 major objections 7 minor 43 references

Unconstrained Body Recognition at Altitude and Range: Comparing Four Approaches

T0 review · 3 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read This paper claims that vision transformers outperform ResNets at identifying people from body shape, and that larger input images, not the transformer architecture, drive the best model's accuracy.

desk verdict Useful empirical comparison with a real confound: the 'equivalently-trained' claim breaks because only the transformers get an extra BRIAR fine-tuning stage. read the letter →

arxiv 2502.07130 v1 pith:7QJKMYIY submitted 2025-02-10 cs.CV cs.AIcs.LG

classification cs.CVcs.AIcs.LG
keywords bodyshapeidentificationpersonre-identificationvisiontransformerSwinResNetclothingchangeunconstrainedrecognitionBRIARdataset
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

This paper compares four deep-learning models for long-term person identification from body shape: two transformer-based models (BIDDS and Swin-BIDDS) and two ResNet-based models (LCRIM and NLCRIM), all trained on a shared pool of nearly two million images spanning nine databases. It claims that the transformer models are consistently more accurate than the ResNet models on standard re-identification benchmarks and on the unconstrained BRIAR test set, which includes probes at distances up to 1000 meters, from UAV altitudes, and with clothing changes. The best model, Swin-BIDDS, also exceeds previously published state-of-the-art results on the MARS, MSMT17, and DeepChange benchmarks. Through ablation experiments, the paper argues that Swin-BIDDS's edge over BIDDS comes primarily from its larger 384x384 input images, with a smaller contribution from the Swin architecture. If these claims hold, body shape becomes a more useful biometric for recognizing people when faces are too distant, too high, or otherwise unavailable.

What carries the argument

The machinery that carries the argument is a set of four models trained from the same data and loss, together with a two-factor ablation. BIDDS is a ViT-B/16 vision transformer; Swin-BIDDS is a Swin Transformer whose shifted windows make it scale efficiently to larger images; LCRIM and NLCRIM share a ResNet-50 backbone, with LCRIM adding a linguistic pre-training stage that maps images to averaged human body descriptors. The ablation isolates the two differences between Swin-BIDDS and BIDDS—backbone type and input resolution—by comparing Swin-BIDDS(384,384), Swin-BIDDS(224,224), BIDDS(224,384), and BIDDS(224,224). The metric deltas from this 2x2 comparison are what the paper uses to attribute the performance gain to image size rather than to the shifted-window architecture.

What would settle it

Give the two ResNet models the same extra BRIAR training that the transformers receive—the same 384x384 images and the same training schedule—and rerun the benchmark and BRIAR comparisons; if the ResNets then match or exceed the transformers, the claim that transformer backbones are superior for body identification would be refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that, when trained on the same large-scale dataset of 1.9 million images with the same hard-triplet objective, vision transformer backbones learn body-shape representations that identify people more accurately than ResNet backbones, both on standard re-identification benchmarks and on the BRIAR test set with probes at up to 1000 meters, from UAV altitudes, and with clothing changes. Among the four models, Swin-BIDDS is the most accurate on nearly every metric and every partition, including face-restricted, long-range, and UAV probes. In ablation experiments that vary the backbone (ViT vs Swin-ViT) and input size (224x224 vs 384x384) independently, increasing the image size improves all metrics by a larger margin than changing the architecture, so the paper concludes that image size is the critical factor in Swin-BIDDS's advantage. The paper also reports that adding linguistic body-description pre-training to a ResNet (LCRIM versus NLCRIM) yields only a small improvement.

Load-bearing premise

The comparison assumes that all four models are trained on an equal footing, but only the two transformer models receive extra training on the BRIAR data afterwards, while the two ResNet models do not.

Editorial extensions

If this is right

  • If the paper is right, Swin-BIDDS is the most accurate of these models for long-term body identification, including on face-restricted, long-range (up to 1000 m), and UAV-captured probes.
  • Vision transformer backbones should be preferred over ResNets for body identification when training data and compute allow, since the transformer advantage holds across all four benchmarks and the unconstrained test set.
  • Input image resolution is the dominant factor: models trained or fine-tuned at 384x384 consistently outperform their 224x224 counterparts, so resolution should be increased before changing backbone.
  • The small gap between LCRIM and NLCRIM suggests that linguistic pre-training adds little for this task, so the complementary-information benefit seen in earlier work does not translate to a large standalone accuracy gain here.
  • Clothing-change robustness is achievable: the strong DeepChange results indicate that the models rely on body shape and other clothing-invariant cues, not on apparel.

Reading between the lines

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

  • A fairer test of the 'transformers are superior' claim would give the ResNet models the same extra BRIAR training that the transformers receive; until that is done, part of the architecture gap could be due to the extra training rather than the backbone.
  • If input resolution is truly the critical factor, then testing Swin-BIDDS at still higher resolutions (for example 512x512) on the BRIAR long-range and UAV partitions could reveal whether the gains continue or saturate; the paper does not run this experiment.
  • The paper's use of averaged embeddings over video frames as gallery and probe templates suggests the method is compatible with video-to-video identification; a natural extension would be to fuse body and face embeddings, since face-restricted probes still achieve a rank-1 of about 0.34 with Swin-BIDDS, leaving room for complementary cues.
  • Because the training pool mixes datasets with and without clothing changes, the relative contribution of clothing-change data to the observed robustness is not isolated; an ablation that removes the clothes-change datasets from training would test whether the DeepChange and BRIAR gains come from those specific training examples.
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

3 major / 7 minor

Summary. The paper presents an empirical comparison of four models for long-term person identification from body shape: BIDDS (ViT-B/16), Swin-BIDDS (Swin-ViT), LCRIM (ResNet-50 with linguistic pre-training), and NLCRIM (ResNet-50 without linguistic pre-training). All models are trained on a common large dataset of ~1.9M images across 9 databases, then evaluated on four standard re-ID benchmarks (MARS, MSMT17, Outdoor Gait, DeepChange) and on the unconstrained BRIAR test set, including altitude, long-range, and clothing-change partitions. The paper claims that Vision Transformers are superior to equivalently-trained ResNets, that Swin-BIDDS is the best model overall, that linguistic pre-training gives only a small advantage, and that input image size is the critical factor behind Swin-BIDDS's advantage over BIDDS. Ablation experiments vary backbone (ViT vs Swin) and input resolution to separate these factors.

Significance. If the central claims hold, the paper would provide a useful, large-scale comparison of backbone architectures for the understudied task of long-term body identification in unconstrained conditions, with direct relevance to the BRIAR program. The paper's strengths include the scale and diversity of the training data, evaluation on an externally defined unconstrained test set with meaningful partitions, and a clean ablation design that separates architecture and image size within the transformer family. The result that image size, rather than architecture alone, drives much of the performance gain is a concrete and falsifiable finding. However, as detailed below, the headline 'equivalently-trained' comparison is not supported by the described training protocols, and several claims rely on single splits with no uncertainty quantification, so the significance of the architecture-level conclusion is currently limited.

major comments (3)
  1. [I and II.C.2] The first claimed contribution is that 'ViT models are superior to equivalently-trained ResNet models for body identification,' but the training protocols in Sections II.A.1 and II.C.2 are not equivalent. After the shared large-scale training on Table I, BIDDS and Swin-BIDDS receive an additional fine-tuning stage on BRS1–5 at 384×384 input resolution, while LCRIM and NLCRIM receive no such fine-tuning stage. Because BRS1–5 is already part of the shared Table I training set, the transformer models effectively see those 697,348 images twice and at a higher input resolution. This confounds architecture with training schedule, data exposure, and input resolution. Tables II and IV therefore do not support the stated equivalence claim, and the paper's central narrative that transformers are inherently better for body identification is weakened. The authors should either train ResNet baselines with the same two-stage fine-tuning protocol and 384×384 input, or explicitly rephrase the claim to describe the specific training protocols used.
  2. [III.A] The DeepChange state-of-the-art claim in Section III.A.2 ('Swin-BIDDS exceeded the SOTA for DeepChange (previous Rank 1 SOTA (.48) [35])') is based on a modified evaluation protocol. Section III.A.1 states that the authors 'restructured the partitioning of the probe and gallery sets' so that probe and gallery use different clothing sets, whereas the original DeepChange benchmark uses similar clothing for each identity across probe and gallery. This is a different protocol, so the comparison to the previously published SOTA number is not valid unless results are also reported under the original protocol. The same concern applies to the MARS and MSMT17 SOTA comparisons, for which the paper does not specify whether the evaluation protocol matches the original benchmark exactly.
  3. [Tables II-IV] All reported results are point estimates from a single training run and a single split, with no error bars, confidence intervals, or significance tests. Claims of consistent superiority across models (e.g., 'Swin-BIDDS performed best on all metrics and for all datasets') are strong and would benefit from at least bootstrap resampling across identities or paired statistical tests, especially for the small UAV partition in Table III (139 identities, 834 media files). Without uncertainty quantification, it is difficult to assess whether the observed differences, particularly the smaller differences in Table V between BIDDS and Swin-BIDDS at matched image size, are reliable.
minor comments (7)
  1. [Table I] The DeepChange training image count appears as '28,1731', which is likely a typo for '281,731' or another value; please check and correct.
  2. [I] The reference list in the second paragraph of the Introduction contains duplicated citations: '[1], [7], [14], [16], [22], [23], [26], [36], [38], [26], [23], [16]' repeats [26], [23], and [16]. Please clean this list.
  3. [References] Reference [31] is listed as 'Outdoor Gait' but the cited title is 'Learning to see people like people: Predicting social impressions of faces,' which does not appear to be the Outdoor Gait dataset paper. Please verify and correct this citation.
  4. [II.C.2] The sentence 'Note: this training data was included in the large scale training and repeated in the fine-tune stage' is an honest disclosure, but it directly supports the major concern about the equivalently-trained comparison; the authors should address this explicitly in the text rather than leaving it as an aside.
  5. [Figure 3] The caption uses '224 px 2, 384 px 2', which should likely be '224×224 px, 384×384 px'; the superscript formatting appears to have been lost.
  6. [Ethical Impact] The ethical impact statement contains the placeholder 'XXXX' for the funding agency name; this should be completed before publication.
  7. [Throughout] There are several typographical inconsistencies, including 'UA V' instead of 'UAV' in the abstract, and 'BIDDs' instead of 'BIDDS' in Section III.A.2. A careful proofreading pass is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical comparison with external benchmarks; disclosed training differences are a validity concern, not a reduction.

full rationale

This paper is an empirical benchmark comparison, not a derivation, so the circularity patterns that involve fitted parameters masquerading as predictions or self-citations forcing conclusions do not apply. The main claims (ViT models outperform ResNet models; Swin-BIDDS is best; image size is the largest contributor) are supported by measurements on external test sets (MARS, MSMT17, Outdoor Gait, DeepChange, and the BRIAR Test Set) that are not used for fitting. The ablation in Section III.C separates architecture from input size with matched comparison models, and the differences are reported as arithmetic differences of measured metrics, not as quantities defined by the model. The only near-issue is that Sections II.A.1 and II.C.2 show BIDDS and Swin-BIDDS receive an additional BRIAR fine-tuning stage while the ResNet models do not, which undermines the 'equivalently-trained' wording as a fairness matter. That is a confound in the experimental comparison, not a definitional equivalence or fitted-input circularity: the reported numbers are not true by construction, and a ResNet trained with the same schedule could in principle close the gap. Self-citations to the authors' earlier LCRIM/NLCRIM paper [26] define the baselines, but those baselines are re-implemented with described changes and evaluated against external benchmarks, so the citation is not load-bearing in a way that forces any result. No circular step can be exhibited from the paper's own equations or definitions.

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

The paper is an empirical comparison, so there are no analytic free parameters or invented entities. The central results rest on assumptions about dataset splits, template-based evaluation protocols, and the validity of BTS partitions; none are independently verified here.

assumptions (4)
  • domain assumption Test identities in MSMT17, MARS, and DeepChange are disjoint from training identities, making the withheld test splits valid.
    Stated in Section II.C as 'a subset of test data were withheld', but identity disjointness is assumed rather than demonstrated.
  • domain assumption Gallery embeddings formed by averaging still-image embeddings and probe embeddings formed by averaging frame embeddings constitute a valid identification protocol.
    Section III.A.1 and III.B.1 describe template averaging; no validation against a standard protocol is provided.
  • domain assumption BRIAR BTS partition labels (face included/restricted, long-range, UAV) are accurate and independent of model selection.
    Section III.B relies on BTS partitions from [3]; no manual audit is reported.
  • domain assumption ImageNet-1k pretrained weights provide a useful initialization for all backbones.
    All models use ImageNet pretraining (Section II.A); this is standard practice but an unstated assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unconstrained Body Recognition at Altitude and Range: Comparing Four Approaches." pith.science (2026). https://pith.science/paper/7QJKMYIY

@misc{pith2026250207130,
  author       = {Pith},
  title        = {Pith review of: Unconstrained Body Recognition at Altitude and Range: Comparing Four Approaches},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7QJKMYIY}},
  note         = {Machine review of arXiv:2502.07130}
}
read the original abstract

This study presents an investigation of four distinct approaches to long-term person identification using body shape. Unlike short-term re-identification systems that rely on temporary features (e.g., clothing), we focus on learning persistent body shape characteristics that remain stable over time. We introduce a body identification model based on a Vision Transformer (ViT) (Body Identification from Diverse Datasets, BIDDS) and on a Swin-ViT model (Swin-BIDDS). We also expand on previous approaches based on the Linguistic and Non-linguistic Core ResNet Identity Models (LCRIM and NLCRIM), but with improved training. All models are trained on a large and diverse dataset of over 1.9 million images of approximately 5k identities across 9 databases. Performance was evaluated on standard re-identification benchmark datasets (MARS, MSMT17, Outdoor Gait, DeepChange) and on an unconstrained dataset that includes images at a distance (from close-range to 1000m), at altitude (from an unmanned aerial vehicle, UAV), and with clothing change. A comparative analysis across these models provides insights into how different backbone architectures and input image sizes impact long-term body identification performance across real-world conditions.

Figures

Figures reproduced from arXiv: 2502.07130 by the authors.

Figure 1
Figure 1. Example body images from the BTS dataset [3]. Subject consented [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Ablation: CMC and ROC curves for the architecture and image size comparisons show that image size is the critical factor in the superior [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Difference in performance by model architecture (ViT, Swin-ViT) and input image size (224 px [PITH_FULL_IMAGE:figures/full_fig_p007_3.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. [35]

    Xu and X

    P. Xu and X. Zhu. Deepchange: A long-term person re-identification benchmark with clothes change. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 11196–11205, 2023

  2. [1]

    J. Chen, X. Jiang, F. Wang, J. Zhang, F. Zheng, X. Sun, and W.- S. Zheng. Learning 3d shape feature for texture-insensitive person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 8146–8155, 2021

  3. [2]

    W. Chen, X. Xu, J. Jia, H. Luo, Y . Wang, F. Wang, R. Jin, and X. Sun. Beyond appearance: a semantic controllable self-supervised learning framework for human-centric visual tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15050–15061, 2023

  4. [3]

    Cornett, J

    D. Cornett, J. Brogan, N. Barber, D. Aykac, S. Baird, N. Burchfield, C. Dukes, A. Duncan, R. Ferrell, J. Goddard, et al. Expanding accurate person recognition to new altitudes and ranges: The briar dataset. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 593–602, 2023

  5. [4]

    Corona, K

    K. Corona, K. Osterdahl, R. Collins, and A. Hoogs. Meva: A large- scale multiview, multimodal video dataset for activity detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1060–1068, January 2021

  6. [5]

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4690– 4699, 2019

  7. [6]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 , 2020

  8. [7]

    X. Gu, H. Chang, B. Ma, S. Bai, S. Shan, and X. Chen. Clothes- changing person re-identification with rgb modality only. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1060–1069, 2022

Show all 43 references
  1. [8]

    C. A. Hahn, A. J. O’Toole, and P. J. Phillips. Dissecting the time course of person recognition in natural viewing environments. British Journal of Psychology , 107(1):117–134, 2016

  2. [9]

    K. Han, S. Gong, Y . Huang, L. Wang, and T. Tan. Clothing-change feature augmentation for person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22066–22075, 2023

  3. [10]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition, 2015

  4. [11]

    T. He, X. Jin, X. Shen, J. Huang, Z. Chen, and X.-S. Hua. Dense interaction learning for video-based person re-identification. In Pro- ceedings of the IEEE/CVF International Conference on Computer Vision, pages 1490–1501, 2021

  5. [12]

    Hermans, L

    A. Hermans, L. Beyer, and B. Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 , 2017

  6. [13]

    M. Q. Hill, S. Streuber, C. A. Hahn, M. J. Black, and A. J. O’Toole. Creating body shapes from verbal descriptions by linking similarity spaces. Psychological science, 27(11):1486–1497, 2016

  7. [14]

    P. Hong, T. Wu, A. Wu, X. Han, and W.-S. Zheng. Fine-grained shape- appearance mutual learning for cloth-changing person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10513–10522, 2021

  8. [15]

    Huang, R

    S. Huang, R. P. Kathirvel, Y . Guo, C. P. Lau, and R. Chellappa. Whole- body detection, identification and recognition at altitude and range. IEEE Transactions on Biometrics, Behavior, and Identity Science , 2024

  9. [16]

    Huang, R

    S. Huang, R. P. Kathirvel, C. P. Lau, and R. Chellappa. Whole-body detection, recognition and identification at altitude and range. arXiv preprint arXiv:2311.05725, 2023

  10. [17]

    Huang, R

    S. Huang, R. Prabhakar, Y . Guo, R. Chellappa, and C. Peng. Vills: Video-image learning to learn semantics for person re-identification, 2024

  11. [18]

    Huang, Y

    S. Huang, Y . Zhou, R. P. Kathirvel, R. Chellappa, and C. P. Lau. Self-supervised learning of whole and component-based semantic representations for person re-identification, 2023

  12. [19]

    M. Kim, A. K. Jain, and X. Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 18750–18759, 2022

  13. [20]

    S. A. Kumar, E. Yaghoubi, A. Das, B. Harish, and H. Proenc ¸a. The p-destre: A fully annotated dataset for pedestrian detection, tracking, and short/long-term re-identification from aerial devices. IEEE Transactions on Information Forensics and Security , 16:1696– 1708, 2020

  14. [21]

    T. Li, J. Liu, W. Zhang, Y . Ni, W. Wang, and Z. Li. Uav-human: A large benchmark for human behavior understanding with unmanned aerial vehicles. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 16261–16270, 2021

  15. [22]

    F. Liu, M. Kim, Z. Gu, A. Jain, and X. Liu. Learning clothing and pose invariant 3d shape representation for long-term person re- identification. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision , pages 19617–19626, 2023

  16. [23]

    F. Liu, M. Kim, Z. Ren, and X. Liu. Distilling clip with dual guidance for learning discriminative human body shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 256–266, 2024

  17. [24]

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision , pages 10012–10022, 2021

  18. [25]

    Loper, N

    M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black. Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG), 34(6):1–16, 2015

  19. [26]

    B. A. Myers, L. Jaggernauth, T. M. Metz, M. Q. Hill, V . N. Gandi, C. D. Castillo, and A. J. O’Toole. Recognizing people by body shape using deep networks of images and words. Proceedings of the IEEE: International Joint Conference on Biometrics , 2023

  20. [27]

    A. J. O’Toole, J. Harms, S. L. Snow, D. R. Hurst, M. R. Pappas, J. H. Ayyad, and H. Abdi. A video database of moving faces and people. IEEE Transactions on pattern analysis and machine intelligence, 27(5):812–816, 2005

  21. [28]

    X. Qian, W. Wang, L. Zhang, F. Zhu, Y . Fu, T. Xiang, Y .-G. Jiang, and X. Xue. Long-term cloth-changing person re-identification. In Proceedings of the Asian Conference on Computer Vision , 2020

  22. [29]

    A. Rice, P. J. Phillips, and A. O’Toole. The role of the face and body in unfamiliar person identification. Applied Cognitive Psychology , 27(6):761–768, 2013

  23. [30]

    Russakovsky, J

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211–252, 2015

  24. [31]

    A. Song, L. Linjie, C. Atalla, and G. Gottrell. Learning to see people like people: Predicting social impressions of faces. Cognitive Science, 2017

  25. [32]

    Streuber, M

    S. Streuber, M. A. Quiros-Ramirez, M. Q. Hill, C. A. Hahn, S. Zuffi, A. O’Toole, and M. J. Black. Body talk: Crowdshaping realistic 3d avatars with words. ACM Transactions on Graphics (TOG) , 35(4):1– 14, 2016

  26. [33]

    H. Wang, Y . Wang, Z. Zhou, X. Ji, D. Gong, J. Zhou, Z. Li, and W. Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265–5274, 2018

  27. [34]

    L. Wei, S. Zhang, W. Gao, and Q. Tian. Person transfer gan to bridge domain gap for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 79–88, 2018

  28. [36]

    Q. Yang, A. Wu, and W.-S. Zheng. Person re-identification by contour sketch under moderate clothing change. IEEE transactions on pattern analysis and machine intelligence , 43(6):2029–2046, 2019

  29. [37]

    Z. Yang, M. Lin, X. Zhong, Y . Wu, and Z. Wang. Good is bad: Causal- ity inspired cloth-debiasing for cloth-changing person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1472–1481, 2023

  30. [38]

    M. Ye, J. Shen, G. Lin, T. Xiang, L. Shao, and S. C. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence , 44(6):2872–2893, 2021

  31. [39]

    M. Ye, J. Shen, G. Lin, T. Xiang, L. Shao, and S. C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Transactions on Pattern Analysis and Machine Intelligence , 44(6):2872–2893, 2022

  32. [40]

    D. Yi, Z. Lei, S. Liao, and S. Z. Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923 , 2014

  33. [41]

    Yovel and A

    G. Yovel and A. J. O’Toole. Recognizing people in motion. Trends in cognitive sciences , 20(5):383–395, 2016

  34. [42]

    Zheng, Z

    L. Zheng, Z. Bie, Y . Sun, J. Wang, C. Su, S. Wang, and Q. Tian. Mars: A video benchmark for large-scale person re-identification. In Computer Vision–ECCV 2016: 14th European Conference, Amster- dam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 868–884....

  35. [43]

    Zheng, L

    L. Zheng, L. Shen, L. Tian, S. Wang, J. Wang, and Q. Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision , pages 1116–1124, 2015

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.