Pith. sign in

REVIEW 4 major objections 4 minor 93 references

Colors See Colors Ignore: Clothes Changing ReID with Color Disentanglement

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Color histograms learned by a dedicated transformer token can act as an annotation-free proxy for clothing; disentangling them from identity features via two-step self-attention improves clothes-changing person re-identification on four…

desk verdict A useful incremental idea undermined by an inconsistent loss definition and optimistic hyperparameter reporting; worth a serious look if the equations are fixed. read the letter →

arxiv 2507.07230 v1 pith:YMMA62XT submitted 2025-07-09 cs.CV

Priyank Pathak , Yogesh S. Rawat This is my paper · ORCID
classification cs.CV
keywords clothes-changingpersonre-identificationcolordisentanglementhistogramself-attentiontransformerappearancebiasRGB-onlyvideo
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

Clothes-changing person re-identification must recognize people when their clothing changes, which makes clothing color a misleading cue. This paper claims that color, used as a learnable proxy rather than a label, can absorb exactly that misleading appearance bias: a dedicated color token predicts color histograms from raw RGB input while a two-step self-attention and an orthogonality loss keep the identity features free of color information. The authors report that this RGB-only scheme raises top-1 accuracy over their strong baseline by 2.9% and 5.0% on the LTCC and PRCC image benchmarks and by 1.0% and 2.5% on the CCVID and MeVID video benchmarks, all without clothing annotations, text attributes, or extra modalities. A sympathetic reader would care because it offers a parameter-lean route to a practical surveillance problem where labeled clothing attributes are expensive and short-lived.

What carries the argument

The load-bearing mechanism is a Color token, a second learnable head token that runs alongside the ReID class token through every transformer block and is trained to predict flattened color histograms from raw RGB input. The second piece is S2A self-attention: two independent self-attention computations per block, one that removes the Color token and one that removes the ReID token, with the spatial tokens' attention averaged across the two passes, which prevents the ReID and Color tokens from directly exchanging information while keeping a single shared transformer. The third piece is the loss $L_{DE}$ that maximizes the cosine distance between $\mathbf{f}_{CO}$ and $\mathbf{f}_{ReID}$, pushing the two embeddings toward orthogonality in the shared feature space.

What would settle it

Train CSCI on a partition of PRCC or LTCC that excludes all clothing colors present in the test set; if top-1 accuracy on the CC protocol falls back to the baseline's level, the model was using memorized color cues rather than a generalizable disentanglement. Alternatively, fit a linear classifier to the color embedding $\mathbf{f}_{CO}$ to predict identity; if it clearly exceeds chance, identity information has leaked past the disentanglement loss.

Watch

Extended reading notes

Core claim

The paper's central claim is that color—specifically the histogram of foreground and background colors in an image or video frame—is a sufficient and cheap proxy for the clothing-driven appearance bias that hurts clothes-changing person re-identification. To use it, CSCI appends a Color token alongside the standard ReID class token in a vision transformer, and trains that token to regress flattened color histograms (either 3D pixel binning or RGB-uv log-chroma projections). Two mechanisms keep the color signal from contaminating identity: S2A self-attention, which runs separate self-attention passes for the ReID and Color tokens and averages the spatial-token results, and a disentanglement loss that maximizes the cosine distance between the two embeddings. The paper argues this is better than full token mixing, better than simply masking ReID–Color attention, and better than using clothing labels as a token, because colors adapt to illumination and background without needing annotation.

Load-bearing premise

The design assumes that pushing the color embedding and the identity embedding apart in cosine space removes clothing-related appearance bias while leaving identity cues intact—that is, color and identity are linearly separable in the shared transformer feature space, and the paper's own cluster analysis shows faces and poses grouped inside the color embedding, indicating the separation is incomplete.

Editorial extensions

If this is right

  • Color embeddings learned without any annotation lift top-1 accuracy by 2.9% on LTCC and 5.0% on PRCC over the EVA-02 image baseline, and by 1.0% on CCVID and 2.5% on MeVID over the EZ-CLIP video baseline.
  • S2A self-attention beats both traditional self-attention and masked self-attention on the CC protocol, with only a 4.39% FLOP increase across 24 blocks.
  • The Color token transfers to lighter transformer backbones (ViT-S and ViT-B), improving TMGF by about 7% top-1 on LTCC.
  • Learned color embeddings correlate with ground-truth clothing labels even though no clothing label is used at training time, supporting color as an annotation-free proxy.
  • Feeding raw color histograms directly is worse than learning to predict them, and test-time color dependency is avoided because only the ReID token is used at inference.

Reading between the lines

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

  • The paper's own cluster analysis shows faces and poses leaking into the color embedding, so a fully separate color branch—at higher compute cost—might disentangle further; this is a natural next test rather than a claim the paper makes.
  • If color is a viable clothing proxy here, the same Color-token recipe could be applied to other appearance-bias problems, such as vehicle re-identification or general object re-ID, where clothing labels do not exist.
  • RGB-uv projections outperformed pixel binning on most benchmarks, suggesting that illumination-robust color representations matter more as scenes get harder; one testable extension is to combine both views or add a color-augmentation scheme.
  • Because the method adds no supervision, it could be plugged into self-supervised or unsupervised ReID pipelines that currently rely on pseudo-labels for appearance debiasing.
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

4 major / 4 minor

Summary. The paper proposes CSCI (Colors See, Colors Ignore), an RGB-only method for clothes-changing person re-identification. The method adds a learnable Color token to a ViT (EVA-02) encoder, computes a color histogram (pixel binning or RGB-uv) from the input image as a self-supervised target, trains the Color token to regress this histogram, and applies an orthogonality loss between the Color token and the ReID token to reduce clothing-related appearance bias. A two-step S2A self-attention is introduced to limit information exchange between the two tokens. Experiments on LTCC, PRCC, CCVID, and MeVID report improvements over the EVA-02 baseline, and additional experiments show gains when CSCI is plugged into smaller transformer backbones.

Significance. If the mechanism works as described, CSCI would be a genuinely lightweight alternative to clothing labels or external attribute models: color histograms can be computed on the fly with no annotations, and the method is demonstrated across four datasets and both image and video ReID. The paper also ships useful analysis (t-SNE of color tokens, GradCAM, error categorization) and code. However, the central training objective is stated inconsistently between the main paper and the supplementary, and the empirical evaluation is weakened by best-configuration hyperparameter reporting, lack of error bars, and a backbone that dwarfs most prior work. The paper's own limitation section and Figure 5b admit incomplete disentanglement, which tempers the 'Color Ignore' claim. The idea is promising, but the current manuscript does not yet securely establish the mechanism or the robustness of the reported gains.

major comments (4)
  1. [§3.3 and Supplementary §16, Eq. (19)] The main text (Eq. (7) and surrounding text) states that LColor_MSE is an MSE regression loss on fCO for predicting the color representation vector, but Supplementary Eq. (19) defines LColor_MSE = ||fCO − fReID||². This is not a regression to the color histogram; it pulls the color token toward the ReID token, directly conflicting with the disentanglement loss LDE (Eq. (1)) that pushes them apart. If Eq. (19) reflects the implementation, the color token does not learn color embeddings as claimed, and the clustering analyses in Figs. 5 and 6 would not follow from the stated objective. If Eq. (19) is a typo, the correct regression target is absent from the manuscript, making the method non-reproducible. This is a load-bearing inconsistency and must be resolved in revision.
  2. [Supplementary §11, Table 6 and Fig. 10] The supplementary states that color histogram hyperparameters 'differ such as reporting the best accuracy obtained for a particular experiment' and Table 6 lists different bin sizes, smoothness factors, normalization schemes, and scale factors for each reported result. With only two runs per experiment and no held-out validation split for hyperparameter selection, the reported gains (+2.9% on LTCC, +5.0% on PRCC, etc.) reflect the best of many configurations rather than a prespecified protocol. This makes the empirical claims optimistic. The authors should report mean and standard deviation over multiple runs for a fixed configuration, or perform hyperparameter selection on a validation split and report the corresponding test numbers.
  3. [§5.2, Fig. 5b and Supplementary §19] The paper's own analysis admits a cross-information leak between fReID and fCO: Figure 5b shows clusters in fCO grouped by similar poses and faces, and §19 states that 'Figure 5 hints at a cross-information leak between fReID and fCO.' This directly weakens the central 'Color Ignore' claim, as the disentanglement is incomplete. The observed accuracy gains might arise from the auxiliary color-prediction task acting as a regularizer rather than from genuine removal of color bias. The authors should provide a more direct measure of color-bias removal, for example by evaluating clothing-attribute prediction accuracy or clothing-similarity of nearest neighbors from fReID before and after applying CSCI, or by tempering the claims accordingly.
  4. [Tables 1 and 2] The main comparisons use EVA-02 (303M parameters) for CSCI, while most prior methods are based on ResNet or smaller transformers. The EVA-02 baseline itself already outperforms or matches prior state of the art on several metrics (e.g., PRCC mAP 59.0 vs. 58.3 for CCPG), so the reported 'state-of-the-art' comparisons are not like-for-like. The generalization experiments in Table 3 (TransReID, TMGF, PAT, TCiP) are helpful but are on the same-clothes setting for several models, and the paper should either provide same-backbone comparisons for the main baselines or clearly frame the comparisons as 'system-level' with different compute budgets.
minor comments (4)
  1. [Eq. (1) and Supplementary Eq. (20)] The disentanglement loss uses the absolute value of the cosine similarity between fCO and fReID; the text says the cosine distance is 'maximized.' Minimizing the absolute cosine similarity only corresponds to maximizing cosine distance if the angle is constrained to be at least 90 degrees, which is not stated. Please clarify the exact objective.
  2. [§5.1] There is a typo in the sentence 'random initialized EV A achieves a 2.6% Top-1 accuracy on LTCC (CC)' — 'EV A' should be 'EVA-02' or 'EVA', and the sentence would benefit from a period after '(CC)'.
  3. [Table 3] The table mixes CC-ReID models (TCiP) with same-clothes ReID models (TransReID, PAT, TMGF) under the title 'CSCI on other Transformers'; the caption should explicitly distinguish the settings, since a +7% Top-1 gain on TMGF in the same-clothes protocol is not directly comparable to gains on CC-ReID.
  4. [Supplementary §11.1] Table 6 has a few entries where the smoothness factor is marked as not applicable or missing (e.g., the MEVID RGB-uv row), which is confusing because the smoothness factor is a core hyperparameter for RGB-uv histograms; the table would be clearer if all fields were filled with explicit values.

Circularity Check

1 steps flagged · score 6.0 of 10

Supplementary Eq. 19 defines the color-prediction loss as MSE between the color token and the ReID token, making the claimed color embedding reduce to a copy of the ReID feature by construction.

  1. self definitional [Supplementary §16, Eq. 19; main §3.3, Eq. 7]
    "The regression MSE loss for predicting colors : LColor_MSE = ||fCO − fReID ||²_2 (19) ... For predicting the color representation vector, we use MSE regression loss (LColor_MSE) on fCO."

    The only written formula for the 'color prediction' loss makes the ReID token fReID the regression target, not the color histogram described in §3.2. Minimizing LColor_MSE therefore drives fCO toward fReID, so the 'Color See' token is, by construction, a copy of the identity feature it is supposed to be disentangled from. The paper's validation of fCO as a color embedding (Figs. 5-6) and the 'Color Ignore' conclusion are not derivable from this objective; they rely on an unwritten histogram-regression target. This is a fitted/defined target renamed as a color prediction.

full rationale

The paper is an empirical method paper rather than a formal derivation, so most of its claims rest on measured benchmark improvements and ablations rather than on a theorem chain. Those experiments are self-contained: color histograms are computed from the input RGB images, the ReID and color tokens are trained on the same data, and the reported gains are comparisons against external baselines. The use of EZ-CLIP from the same group is a code-released prior technique used as a baseline and component, not a self-citation invoked to forbid alternatives or to prove uniqueness, so it does not constitute load-bearing circularity. The one genuinely circular-by-construction element is Supplementary Eq. 19, where the MSE loss 'for predicting colors' is written as ||fCO − fReID||², i.e. the color token is trained to match the ReID token rather than the color histogram. If implemented as written, the color embedding reduces to a copy of the identity feature, and the central claim that fCO captures color-related appearance bias while being disentangled from fReID is not supported by the paper's own equations. Since the main text's prose suggests the intended target was the histogram, this is partial rather than total circularity, but the written formulation still makes one of the paper's key 'predictions' equivalent to its input by definition.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The central claim depends on color histograms being an adequate proxy for appearance bias, on the orthogonality loss removing color without removing identity, and on the strong EVA-02 backbone. These are assumptions, not derived facts.

free parameters (3)
  • RGB-uv color histogram hyperparameters (bin size, smoothness tau, concat/mean, normalization, scale factor) = Varies per dataset; e.g., LTCC uses bin=32, tau=0.001, concat, L2, scale=100; PRCC uses bin=32, tau=0.001, concat, L1…
    These are tuned per dataset to report best results, as disclosed in Supp. Sec. 11 and Table 6.
  • Pixel binning histogram hyperparameters (bin size, normalization, scale factor) = e.g., LTCC: bin=20, concat, L2, scale=100; PRCC: bin=20, concat, L1, scale=1000
    Also selected per dataset.
  • Training hyperparameters (learning rate, optimizer, epochs, loss weights) = Not fully reported
    The main text gives batch size and input resolution but omits full schedule, so these are presumably fitted empirically.
assumptions (3)
  • ad hoc to paper Color histograms computed from raw RGB images serve as a sufficient proxy for clothing-related appearance bias.
    The entire method rests on this; the paper argues it via clustering examples but provides no formal support.
  • ad hoc to paper The disentanglement loss LDE (Eq. 1) removes color bias without damaging identity cues.
    Assumes linear separability of color and identity in feature space; paper's own Fig. 5b shows leakage.
  • domain assumption EVA-02 pre-trained weights are a strong and appropriate backbone for ReID.
    The method trains on top of this backbone; without it, performance drops (random init gives 2.6% Top-1 on LTCC).
invented entities (1)
  • Color token
    purpose: A learnable token appended to the transformer to encode color-related appearance information, with S2A self-attention to limit its interaction with the ReID token.
    The token is a network component whose effectiveness is demonstrated only on the paper's own datasets; it does not make an external falsifiable prediction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Colors See Colors Ignore: Clothes Changing ReID with Color Disentanglement." pith.science (2026). https://pith.science/paper/YMMA62XT

@misc{pith2026250707230,
  author       = {Pith},
  title        = {Pith review of: Colors See Colors Ignore: Clothes Changing ReID with Color Disentanglement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YMMA62XT}},
  note         = {Machine review of arXiv:2507.07230}
}
read the original abstract

Clothes-Changing Re-Identification (CC-ReID) aims to recognize individuals across different locations and times, irrespective of clothing. Existing methods often rely on additional models or annotations to learn robust, clothing-invariant features, making them resource-intensive. In contrast, we explore the use of color - specifically foreground and background colors - as a lightweight, annotation-free proxy for mitigating appearance bias in ReID models. We propose Colors See, Colors Ignore (CSCI), an RGB-only method that leverages color information directly from raw images or video frames. CSCI efficiently captures color-related appearance bias ('Color See') while disentangling it from identity-relevant ReID features ('Color Ignore'). To achieve this, we introduce S2A self-attention, a novel self-attention to prevent information leak between color and identity cues within the feature space. Our analysis shows a strong correspondence between learned color embeddings and clothing attributes, validating color as an effective proxy when explicit clothing labels are unavailable. We demonstrate the effectiveness of CSCI on both image and video ReID with extensive experiments on four CC-ReID datasets. We improve the baseline by Top-1 2.9% on LTCC and 5.0% on PRCC for image-based ReID, and 1.0% on CCVID and 2.5% on MeVID for video-based ReID without relying on additional supervision. Our results highlight the potential of color as a cost-effective solution for addressing appearance bias in CC-ReID. Github: https://github.com/ppriyank/ICCV-CSCI-Person-ReID.

Figures

Figures reproduced from arXiv: 2507.07230 by the authors.

Figure 1
Figure 1. Clustering of color representation: Clusters (2 images/cluster, same outline color) generated via RGB-uv projection [2]. (a): Clothing changes in video are clustered differently (b & c): Clustering based on similar clothing & background / illumination. (d): Different clothing but similar overall color profile clustered together. Images from CCVID (a) and LTCC (b, c, d). 22, 36], and (2) fully separate dual-branch [3… view at source ↗
Figure 2
Figure 2. CSCI Framework: Image spatial tokens (yellow) are appended with ReID (fReID), and Color (fCO) tokens. Tokens are passed through transformer blocks (modified self-attention), discarding spatial ones after the last block. fReID (inference) pre￾dicts the identity labels and trains with triplet loss. fCO predicts flattened color histograms and disentangled from fReID. neck. Thus our design primarily focuses on transform… view at source ↗
Figure 3
Figure 3. Self-Attention Variants: Dotted arrows and shades of coloring indicate information exchange via self-attention. For explanation, spatial tokens are grouped as one, skipping their internal self-attention. (a) Traditional self-attention, all tokens attend each other; (b) Masked self-attention is the same as traditional without ReID-Color token attention; (c) S2A self-attention has two independent self￾attention steps,… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Color Representation: Top Pixel Binning: 3D His￾togram (20×20×20); Bottom:RGB-uv 2D Projection/Histogram (3×32×32, upscaled for viewability. Images taken from CCVID. 3.4. Generalization to Video ReID We also demonstrate the generalization of CSCI to Video ReID (CSCI-V)…
Figure 5
Figure 5. Figure 5: Clustering of Color token: Images from PRCC gallery set. (a) Two images per cluster (same outline colors), indicating color embedding clusters based on background & illumination in case of the same clothing. (b) Cluster include similar-looking different clothes (5,6) (…
Figure 6
Figure 6. Figure 6: Color token embeddings (fCO) t-SNE: t-SNE clusters of fCO can be separated into (a) clothes labels (left) indicating a correlation between learned fCO and clothes labels. (b) K-means clusters of fCO differ in clothing i.e. different clothing may share similar colors. F…
Figure 7
Figure 7. Figure 7: CSCI vs Baseline: Column of images are top 5 gallery matches for the query, taken from PRCC dataset (clothes changing), with green (red) indicating correct (wrong) matches. Left & Middle are cases where baseline mistakes when the gallery wears similar clothing (left ha…
Figure 8
Figure 8. Figure 8: CSCI Error analysis: Top-1 mismatches on PRCC. fCO effectively identifies similar colors. However, noisy clusters (7,8) hint at a potential information leak between fReID features and fCO as similar poses (7) and faces (8) (hidden for privacy) are grouped together. Thi…
Figure 9
Figure 9. Figure 9: RGB-uv parameters: All images are resized for better visualization (a) Bin size is visualized using a scale factor of 1000 and smoothness of 0.02. (b) Smoothness Factor τ is calculated for 64 bin size and scale factor 1000. (c) Scale factor is calculated for 64 bin siz…
Figure 10
Figure 10. Figure 10: LTCC performance for various RGB-uv hyper pa￾rameters: The red dotted line indicates the baseline @44.9%. The color-coded text indicates the hyperparameters as (Concate/Mean, normalization, and scaling weight). All plots have τ = 0.001. RoPE [47])2 . The model operate…
Figure 11
Figure 11. Figure 11: EVA-02 self-attention with RoPE Emebeddings Fig￾ure is taken from Fang et al.. T1 .…...….. Tn Temporal Token Video Frame Tokens + Temporal Position Embedding 24 x EVA-02 Block T1 T1 Avg Tn Tn Avg Self-attention Next Block Next Block Weight Sharing Self-attention Tn Se…
Figure 12
Figure 12. Figure 12: EZ-CLIP (Video ReID): EVA-02 Video architecture. of spatial tokens is added to a newly introduced trainable temporal token (orange) for each timestamp. These tempo￾ral tokens undergo self-attention using the shared weight of the self-attention layer of spatial tokens.…
Figure 14
Figure 14. Figure 14: CSCI GradCAM (Similar Clothing): Same descrip￾tion as [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

93 extracted references · 71 canonical work pages

  1. [1]

    Association for Computing Machinery

    MORE ’24: Proceedings of the 1st ICMR Workshop on Mul- timedia Object Re-Identification, New York, NY , USA, 2024. Association for Computing Machinery. 18

    2024

  2. [2]

    Sensor-independent illumination estimation for dnn models

    Mahmoud Afifi and Michael S Brown. Sensor-independent illumination estimation for dnn models. In British Machine Vision Conference (BMVC), 2019. 2, 4

    2019

  3. [3]

    Brubaker, and Michael S

    Mahmoud Afifi, Marcus A. Brubaker, and Michael S. Brown. HistoGAN: Controlling colors of GAN-generated and real images via color histograms. In CVPR, 2021. 4, 12

    2021

  4. [4]

    Ez-clip: Efficient zeroshot video action recognition

    Shahzad Ahmad, Sukalpa Chanda, and Yogesh S Rawat. Ez-clip: Efficient zeroshot video action recognition. arXiv preprint arXiv:2312.08010, 2023. 4, 5, 6, 14

    arXiv 2023

  5. [5]

    T2l: Efficient zero-shot action recognition with temporal to- ken learning

    Shahzad Ahmad, Sukalpa Chanda, and Yogesh S Rawat. T2l: Efficient zero-shot action recognition with temporal to- ken learning. Transactions on Machine Learning Research,

  6. [6]

    Geff: improving any clothes-changing person reid model us- ing gallery enrichment with face features

    Daniel Arkushin, Bar Cohen, Shmuel Peleg, and Ohad Fried. Geff: improving any clothes-changing person reid model us- ing gallery enrichment with face features. In Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision, pages 152–162, 2024. 2, 6, 16, 17

    2024

  7. [7]

    Activity- biometrics: Person identification from daily activities

    Shehreen Azad and Yogesh Singh Rawat. Activity- biometrics: Person identification from daily activities. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 287–296, 2024. 2

    2024

  8. [8]

    Disenq: Disentan- gling q-former for activity-biometrics

    Shehreen Azad and Yogesh S Rawat. Disenq: Disentan- gling q-former for activity-biometrics. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025. 2

    2025

Show all 93 references
  1. [9]

    Cloth-changing person re-identification with self-attention

    Vaibhav Bansal, Gian Luca Foresti, and Niki Martinel. Cloth-changing person re-identification with self-attention. In Proceedings of the IEEE/CVF Winter Conference on Ap- plications of Computer Vision (WACV) Workshops , pages 602–610, 2022. 2

    2022

  2. [10]

    Learning 3d shape feature for texture-insensitive person re-identification

    Jiaxing Chen, Xinyang Jiang, Fudong Wang, Jun Zhang, Feng Zheng, Xing Sun, and Wei-Shi Zheng. Learning 3d shape feature for texture-insensitive person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 8146–8155,

  3. [11]

    Beyond appearance: a semantic controllable self-supervised learning framework for human-centric visual tasks

    Weihua Chen, Xianzhe Xu, Jian Jia, Hao Luo, Yaohua Wang, Fan Wang, Rong Jin, and Xiuyu Sun. Beyond appearance: a semantic controllable self-supervised learning framework for human-centric visual tasks. In The IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2023. 2

    2023

  4. [12]

    Rest- reid: Transformer block-based residual learning for person re-identification

    Ying Chen, Shixiong Xia, Jiaqi Zhao, Yong Zhou, Qiang Niu, Rui Yao, Dongjun Zhu, and Dongjingdian Liu. Rest- reid: Transformer block-based residual learning for person re-identification. Pattern Recognition Letters , 157:90–96,

  5. [13]

    Expanding accurate person recognition to new altitudes and ranges: The briar dataset

    David Cornett, Joel Brogan, Nell Barber, Deniz Aykac, Seth Baird, Nicholas Burchfield, Carl Dukes, Andrew Duncan, Regina Ferrell, Jim Goddard, et al. Expanding accurate person recognition to new altitudes and ranges: The briar dataset. In Proceedings of the IEEE/CVF Winter Con...

    2023

  6. [14]

    Dcr-reid: Deep component reconstruction for cloth-changing person re-identification

    Zhenyu Cui, Jiahuan Zhou, Yuxin Peng, Shiliang Zhang, and Yaowei Wang. Dcr-reid: Deep component reconstruction for cloth-changing person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):4415– 4428, 2023. 5, 17

    2023

  7. [15]

    Mevid: Multi-view extended videos with identities for video person re-identification

    Daniel Davila, Dawei Du, Bryon Lewis, Christopher Funk, Joseph Van Pelt, Roderic Collins, Kellie Corona, Matt Brown, Scott McCloskey, Anthony Hoogs, and Brian Clipp. Mevid: Multi-view extended videos with identities for video person re-identification. In IEEE/CVF Winter Confer...

    2023

  8. [16]

    Bert: Pre-training of deep bidirectional trans- formers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional trans- formers for language understanding. In North American Chapter of the Association for Computational Linguistics ,

  9. [17]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint a...

    2010 arXiv

  10. [18]

    Opengait: Revisiting gait recognition towards better practicality

    Chao Fan, Junhao Liang, Chuanfu Shen, Saihui Hou, Yongzhen Huang, and Shiqi Yu. Opengait: Revisiting gait recognition towards better practicality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9707–9716, 2023. 2

    2023

  11. [19]

    Eva-02: A visual representa- tion for neon genesis

    Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xin- long Wang, and Yue Cao. Eva-02: A visual representa- tion for neon genesis. Image and Vision Computing , page 105171, 2024. 3, 5, 13, 15

    2024

  12. [20]

    Clothes-changing person re-identification with rgb modality only

    Xinqian Gu, Hong Chang, Bingpeng Ma, Shutao Bai, Shiguang Shan, and Xilin Chen. Clothes-changing person re-identification with rgb modality only. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1060–1069, 2022. 1, 2, 5, 12, 15, 17, 18

    2022

  13. [21]

    You can run but not hide: Improving gait recognition with intrinsic occlusion type awareness

    Ayush Gupta and Rama Chellappa. You can run but not hide: Improving gait recognition with intrinsic occlusion type awareness. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages 5893–5902, 2024. 1

    2024

  14. [22]

    Clothing-change feature augmentation for person 9 re-identification

    Ke Han, Shaogang Gong, Yan Huang, Liang Wang, and Tie- niu Tan. Clothing-change feature augmentation for person 9 re-identification. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 22066–22075, 2023. 2, 5, 15, 17

    2023

  15. [23]

    Transreid: Transformer-based object re- identification

    Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re- identification. In Proceedings of the IEEE/CVF international conference on computer vision , pages 15013–15022, 2021. 2, 4, 5, 6

    2021

  16. [24]

    Instruct-reid: A multi-purpose person re-identification task with instruc- tions

    Weizhen He, Yiheng Deng, Shixiang Tang, Qihao Chen, Qingsong Xie, Yizhou Wang, Lei Bai, Feng Zhu, Rui Zhao, Wanli Ouyang, Donglian Qi, and Yunfeng Yan. Instruct-reid: A multi-purpose person re-identification task with instruc- tions. In Proceedings of the IEEE/CVF Conference o...

    2024

  17. [25]

    Learning disentangled representation implicitly via trans- former for occluded person re-identification

    Mengxi Jia, Xinhua Cheng, Shijian Lu, and Jian Zhang. Learning disentangled representation implicitly via trans- former for occluded person re-identification. IEEE Trans- actions on Multimedia, 25:1294–1305, 2022. 2

    2022

  18. [26]

    Cloth-changing person re- identification from a single image with gait prediction and regularization

    Xin Jin, Tianyu He, Kecheng Zheng, Zhiheng Yin, Xu Shen, Zhen Huang, Ruoyu Feng, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. Cloth-changing person re- identification from a single image with gait prediction and regularization. In Proceedings of the IEEE/CVF Conference on ...

    2022

  19. [27]

    Transformer based multi-grained features for unsupervised person re- identification

    Jiachen Li, Menglin Wang, and Xiaojin Gong. Transformer based multi-grained features for unsupervised person re- identification. In Proceedings of the IEEE/CVF Winter Con- ference on Applications of Computer Vision (WACV) Work- shops, pages 42–50, 2023. 5, 6

    2023

  20. [28]

    Rethinking clothes changing person reid: Conflicts, synthesis, and optimization

    Junjie Li, Guanshuo Wang, Fufu Yu, Yichao Yan, Qiong Jia, Shouhong Ding, Xingdong Sheng, Yunhui Liu, and Xi- aokang Yang. Rethinking clothes changing person reid: Conflicts, synthesis, and optimization. arXiv preprint arXiv:2404.12611, 2024. 5, 6, 17

    2024 arXiv

  21. [29]

    Dc-former: Di- verse and compact transformer for person re-identification

    Wen Li, Cheng Zou, Meng Wang, Furong Xu, Jianan Zhao, Ruobing Zheng, Yuan Cheng, and Wei Chu. Dc-former: Di- verse and compact transformer for person re-identification. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 1415–1423, 2023. 3

    2023

  22. [30]

    Re- fining video-based person re-identification: An integrated framework with facial and body cues

    Yichen Li, Yufei Yin, Wengang Zhou, and Houqiang Li. Re- fining video-based person re-identification: An integrated framework with facial and body cues. In Proceedings of the 1st ICMR Workshop on Multimedia Object Re-Identification, New York, NY , USA, 2024. Association for Co...

    2024

  23. [31]

    Yu-Jhe Li, Xinshuo Weng, and Kris M. Kitani. Learn- ing shape representations for person re-identification under clothing change. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages 2432–2441, 2021. 7

    2021

  24. [32]

    Differ: Disentangling iden- tity features via semantic cues for clothes-changing person re-id

    Xin Liang and Yogesh S Rawat. Differ: Disentangling iden- tity features via semantic cues for clothes-changing person re-id. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 13980–13989, 2025. 2

    2025

  25. [33]

    Learning clothing and pose invariant 3d shape represen- tation for long-term person re-identification

    Feng Liu, Minchul Kim, ZiAng Gu, Anil Jain, and Xiaoming Liu. Learning clothing and pose invariant 3d shape represen- tation for long-term person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion (ICCV), pages 19617–19626, 2023. 2,...

    2023

  26. [34]

    Distilling clip with dual guidance for learning discrimina- tive human body shape representation

    Feng Liu, Minchul Kim, Zhiyuan Ren, and Xiaoming Liu. Distilling clip with dual guidance for learning discrimina- tive human body shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 256–266, 2024. 1, 2, 5, 6, 15, 17

    2024

  27. [35]

    Learn- ing clothes-irrelevant cues for clothes-changing person re- identification

    Jingyi Mu, Yong Li, Jun Li, and Jian Yang. Learn- ing clothes-irrelevant cues for clothes-changing person re- identification. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022 , page 337. BMV A Press, 2022. 2, 17

    2022

  28. [36]

    Nguyen, Khadija Khaldi, Dung Nguyen, Pranav Mantini, and Shishir Shah

    Vuong D. Nguyen, Khadija Khaldi, Dung Nguyen, Pranav Mantini, and Shishir Shah. Contrastive viewpoint-aware shape learning for long-term person re-identification. InPro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision (WACV), pages 1041–1049, 20...

    2024

  29. [37]

    Nguyen, Pranav Mantini, and Shishir K

    Vuong D. Nguyen, Pranav Mantini, and Shishir K. Shah. Contrastive clothing and pose generation for cloth-changing person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 7541–7549, 2024. 5, 12, 17

    2024

  30. [38]

    Tem- poral 3d shape modeling for video-based cloth-changing per- son re-identification

    Vuong D Nguyen, Pranav Mantini, and Shishir K Shah. Tem- poral 3d shape modeling for video-based cloth-changing per- son re-identification. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , pages 173–182, 2024. 5, 17

    2024

  31. [39]

    Part-aware transformer for generalizable person re- identification

    Hao Ni, Yuke Li, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Part-aware transformer for generalizable person re- identification. In Proceedings of the IEEE/CVF international conference on computer vision , pages 11280–11289, 2023. 2, 5, 6

    2023

  32. [40]

    Fine-grained re-identification

    Priyank Pathak. Fine-grained re-identification. arXiv preprint arXiv:2011.13475, 2020. 1

    2011 arXiv

  33. [41]

    Coarse attribute pre- diction with task agnostic distillation for real world clothes changing reid

    Priyank Pathak and Yogesh S Rawat. Coarse attribute pre- diction with task agnostic distillation for real world clothes changing reid. arXiv preprint arXiv:2505.12580, 2025. 4

    2025

  34. [42]

    Video person re-id: Fantastic techniques and where to find them (student abstract)

    Priyank Pathak, Amir Erfan Eshratifar, and Michael Gormish. Video person re-id: Fantastic techniques and where to find them (student abstract). In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13893– 13894, 2020. 1, 18

    2020

  35. [43]

    Long-term cloth-changing person re-identification

    Xuelin Qian, Wenxuan Wang, Li Zhang, Fangrui Zhu, Yan- wei Fu, Tao Xiang, Yu-Gang Jiang, and Xiangyang Xue. Long-term cloth-changing person re-identification. arXiv preprint arXiv:2005.12633, 2020. 4

    2005 arXiv

  36. [44]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  37. [45]

    Glu variants improve transformer

    Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 13 10

    2002 arXiv

  38. [46]

    Semantic-guided pixel sampling for cloth-changing person re-identification

    Xiujun Shu, Ge Li, Xiao Wang, Weijian Ruan, and Qi Tian. Semantic-guided pixel sampling for cloth-changing person re-identification. IEEE Signal Processing Letters, 28:1365– 1369, 2021. 2

    2021

  39. [47]

    Roformer: Enhanced transformer with rotary position embedding

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063,

  40. [48]

    Training data-efficient image transformers & distillation through at- tention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through at- tention. In International conference on machine learning , pages 10347–10357. PMLR, 2021. 3

    2021

  41. [49]

    When person re-identification meets chang- ing clothes

    Fangbin Wan, Yang Wu, Xuelin Qian, Yixiong Chen, and Yanwei Fu. When person re-identification meets chang- ing clothes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Work- shops, 2020. 1

    2020

  42. [50]

    Foundation transform- ers

    Hongyu Wang, Shuming Ma, Shaohan Huang, Li Dong, Wenhui Wang, Zhiliang Peng, Yu Wu, Payal Bajaj, Sak- sham Singhal, Alon Benhaim, et al. Foundation transform- ers. arXiv preprint arXiv:2210.06423, 2022. 13

    2022 arXiv

  43. [51]

    A benchmark of video-based clothes-changing person re-identification

    Likai Wang, Xiangqun Zhang, Ruize Han, Jialin Yang, Xi- aoyu Li, Wei Feng, and Song Wang. A benchmark of video-based clothes-changing person re-identification. arXiv preprint arXiv:2211.11165, 2022. 6, 17

    2022 arXiv

  44. [52]

    Image-text-image knowledge transferring for lifelong person re-identification with hybrid clothing states

    Qizao Wang, Xuelin Qian, Bin Li, Yanwei Fu, and Xi- angyang Xue. Image-text-image knowledge transferring for lifelong person re-identification with hybrid clothing states. arXiv preprint arXiv:2405.16600, 2024. 2

    2024

  45. [53]

    Exploring fine-grained representation and recom- position for cloth-changing person re-identification

    Qizao Wang, Xuelin Qian, Bin Li, Xiangyang Xue, and Yan- wei Fu. Exploring fine-grained representation and recom- position for cloth-changing person re-identification. IEEE Transactions on Information Forensics and Security , 2024. 2, 5, 6, 17

    2024

  46. [54]

    Pyramid spatial-temporal aggregation for video-based person re-identification

    Yingquan Wang, Pingping Zhang, Shang Gao, Xia Geng, Hu Lu, and Dong Wang. Pyramid spatial-temporal aggregation for video-based person re-identification. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12006–12015, 2021. 18

    2021

  47. [55]

    A transformer-based cloth-irrelevant patches feature ex- tracting method for long-term cloth-changing person re- identification

    Zepeng Wang, Xinghao Jiang, Ke Xu, and Tanfeng Sun. A transformer-based cloth-irrelevant patches feature ex- tracting method for long-term cloth-changing person re- identification. In Advances in Computer Graphics: 39th Computer Graphics International Conference, CGI 2022, Vir...

    2022

  48. [56]

    Adaptive graph representation learn- ing for video person re-identification

    Yiming Wu, Omar El Farouk Bourahla, Xi Li, Fei Wu, Qi Tian, and Xue Zhou. Adaptive graph representation learn- ing for video person re-identification. IEEE Transactions on Image Processing, 29:8821–8830, 2020. 18

    2020

  49. [57]

    Cloth- changing person re-identification with invariant feature pars- ing for uavs applications

    Mingfu Xiong, Xinxin Yang, Hanmei Chen, Wael Hosny Fouad Aly, Abdullah AlTameem, Abdul Khader Jilani Saudagar, Shahid Mumtaz, and Khan Muhammad. Cloth- changing person re-identification with invariant feature pars- ing for uavs applications. IEEE Transactions on Vehicular Tech...

    2024

  50. [58]

    Person re- identification by contour sketch under moderate clothing change

    Qize Yang, Ancong Wu, and Wei-Shi Zheng. Person re- identification by contour sketch under moderate clothing change. IEEE transactions on pattern analysis and machine intelligence, 43(6):2029–2046, 2019. 4

    2019

  51. [59]

    Good is bad: Causality inspired cloth-debiasing for cloth-changing person re-identification

    Zhengwei Yang, Meng Lin, Xian Zhong, Yu Wu, and Zheng Wang. Good is bad: Causality inspired cloth-debiasing for cloth-changing person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1472–1481, 2023. 2, 4, 5,...

    2023

  52. [60]

    Transformer for object re- identification: A survey

    Mang Ye, Shuoyi Chen, Chenyue Li, Wei-Shi Zheng, David Crandall, and Bo Du. Transformer for object re- identification: A survey. arXiv preprint arXiv:2401.06960,

    arXiv

  53. [61]

    Sharc: Shape and appearance recognition for person identification in-the-wild

    Haidong Zhu, Wanrong Zheng, Zhaoheng Zheng, and Ram Nevatia. Sharc: Shape and appearance recognition for person identification in-the-wild. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6290–6300, 2024. 5, 17, 18 11 Colors Se...

    2024

  54. [62]

    (Section 8)

    Intuition behind using colors justification of why colors can be a good lightweight substitute for clothing labels in reducing appearance bias. (Section 8)

  55. [63]

    Description of two Zips attached with this supplemen- tary, the RGB GIFs denoting the clusters based on color representation (Section 9)

  56. [64]

    RGB-uv Histogram Mathematical description (Sec- tion 10)

  57. [65]

    Color Hyperparamters All color profiles related hyperparameters used in the main paper (Section 11)

  58. [66]

    Image Encoder EV A-02 All color profiles related hyperparameters used in the main paper (Section 12)

  59. [67]

    EZ-CLIP Video ReID model architecture (Section 13)

  60. [68]

    Intutive explanation of S2A self-attention (Sec- tion 14)

  61. [69]

    (Section 15)

    Gram CAM Comparison Grad cam visualization for CSCI and comparison with baseline. (Section 15)

  62. [70]

    (Section 16)

    All Losses Formula for all the loss function used to train models. (Section 16)

  63. [71]

    All Results Table include technique based previous works which can are generally applied on pre-trained models (Section 17)

  64. [72]

    (Section 18)

    Mevid Results All Mevid evaluation protocols, namely overall, clothing-based, distance/location-based, and scale-based. (Section 18)

  65. [73]

    Limitation and ethical consideration (Section 19)

  66. [74]

    cloth 070 id 1

    Intuition behind using colors Colors’s inherently have limitations like lacking global context with noisy/ambiguity. However, our goal is not to solve/replace clothing, but rather to offer a lightweight/practical alternative to costly clothing la- bels/annotation. Notably, exi...

  67. [75]

    We used the DBSCAN algorithm to cluster color histograms since we don’t know the number of actual clusters

    RGB GIFs (Figure 1) Figure 1 main submission shows CCVID video frames and LTCC images clustered using RGB-uv projections. We used the DBSCAN algorithm to cluster color histograms since we don’t know the number of actual clusters. In the main paper, we have shown 7 clusters, wh...

  68. [76]

    RGB-uv Histograms derivation (Under Color Representation in Methodology) HistGAN [3] used RGB-uv projection of colors for style transfer in GANs. They implied this representation (2D rep- resentation) has better invariance to illumination, is more compact than RGB pixel binnin...

  69. [77]

    These hyperparameters are described below with ones used in experimental values reported in the main submission listed in Table 6

    Color Histograms hyperparameters(Under Implementation in Experiments) Color histogram hyperparameters differ such as reporting the best accuracy obtained for a particular experiment. These hyperparameters are described below with ones used in experimental values reported in th...

  70. [78]

    h = 32 is shown for all demonstration purposes in the main submission

    Bin size h = u = v: size of histogram images. h = 32 is shown for all demonstration purposes in the main submission

  71. [79]

    Smoothness factor τ: τ = 0.02 is used for all demon- stration purposes in the main submission

  72. [80]

    We flat- ten these histograms to regress over via color tokens

    Concatenating vs Averaging the 3 channels . We flat- ten these histograms to regress over via color tokens. Hence we can either concatenate these channels by cre- ating R3×h×h 1D vector or averaging Rh×h vectors

  73. [81]

    It can be either L2 normal- ized or L1 normalized or simple min, max normalization x−min(x) max(x)−min(x)

    Normalization: Normalized flattened histogram vec- tor is fed to the model. It can be either L2 normal- ized or L1 normalized or simple min, max normalization x−min(x) max(x)−min(x)

  74. [82]

    Once normalized most of the values of histograms are close to 0, e.g

    Scale Multiplication factor. Once normalized most of the values of histograms are close to 0, e.g. 3 × 32 × 32 L2 normalized vector. Hence, we multiply the vectors by 1, 10, 100, or 1000 as a scale factor. These hyperparameters namely, bin size, smoothness factor, and scale fa...

  75. [83]

    h = 20 is constant for all the experiments, producing 20 × 20 × 20 3D his- togram

    Bin size h: size of histogram bins. h = 20 is constant for all the experiments, producing 20 × 20 × 20 3D his- togram

  76. [84]

    Normalization, same as RGB-uv histograms

  77. [85]

    Scale Multiplication factor same as RGB-uv his- tograms. 11.3. Color Hyper parameter performance As shown in Tab. 6, there are several hyperparameters. In order to give some estimate of the relative performance of each hyperparameters, we have plotted LTCC performance for some...

  78. [86]

    It has shown remark- able generalization with state-of-the-art zero-shot image classification capabilities

    Visual Transformer Encoder EV A-02 De- tails (Under Color See Color Ignore (CSCI) in Methodology) EV A-02 [19] is a vision-language transformer with large- scale CLIP based pre-training [44]. It has shown remark- able generalization with state-of-the-art zero-shot image classi...

  79. [87]

    We first pre-train the EV A-02 image encoder on the video dataset, using random frames for Image ReID

    Video ReID : EZ-CLIP : (EC) (Under Gen- eralization to Video ReID in Methodology) Recently [4] proposed using temporal prompts (tokens) for converting any image-based transformers into a video model3. We first pre-train the EV A-02 image encoder on the video dataset, using ran...

  80. [88]

    Traditional

    Intutive Explanation of S2A self-attention S2A is the same as masked self-attention, with one key dif- ference: ID and Color token do not influence each other’s weights. This subtle yet important difference yields 1% gain over masked self-attention (Tab.4(a)), noteworthy as Pe...

  81. [89]

    (a) Different clothing: Figure 13 shows the attention of baseline (vanilla model) and our CSCI, where both models 15 Gallery Vanilla CSCI Query Figure 13

    CSCI GradCAM Grad-CAM is applied on the PRCC train set image (denoted as ‘Query’ and ‘Gallery’) via similarity from two images (query), with and without the similar clothing. (a) Different clothing: Figure 13 shows the attention of baseline (vanilla model) and our CSCI, where ...

  82. [90]

    Triplet loss ( LT ripelt) for a minibatch terms the i-th as ”anchor” (i,A) f iA ReID

    Loss formulations Cross entropy loss for a batch of ‘N’ images, and classifier layer Wk : LID CE = − NX i=1 log eWyi xi+byi PC k=1 eWkxi+bk (17) where yi is the label vector for xi and C classes. Triplet loss ( LT ripelt) for a minibatch terms the i-th as ”anchor” (i,A) f iA R...

  83. [91]

    We have added them here for completion

    All Results (Under Comparison with SOTA Methods in Experiments) In the main paper, technique-based approaches which take existing models and improve their accuracy, mostly via fa- cial features were skipped. We have added them here for completion. We have skipped the results f...

  84. [92]

    MEVID All Protocols ( Under Experiment Settings) While existing works have reported accuracy mostly on the Overall category and Clothes changing category (CC), Original work [15], evaluated baseline models on various metrics as well. These metrics include: (i) SC: Query- 17 Me...

  85. [93]

    To address this, color embeddings might need to be entirely separated from the ReID model, warranting a deeper analysis of computation overhead

    Limitations and Ethical Statement Figure 5 hints at a cross-information leak between fReID and fCO. To address this, color embeddings might need to be entirely separated from the ReID model, warranting a deeper analysis of computation overhead. Furthermore, the color aspect of...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.