Pith. sign in

REVIEW 4 major objections 6 minor 79 references

Generalizable Object Re-Identification via Visual In-Context Prompting

T0 review · 4 major / 6 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 One re-identification model, trained once on a few categories, can match instances of entirely new categories with no retraining: a language model reads a few example pairs and writes visual prompts for a frozen vision transformer.

desk verdict VICP is a solid few-shot ReID framework with a valuable new benchmark, but the paper never proves the LLM is doing the semantic reasoning it claims. read the letter →

arxiv 2508.21222 v1 pith:UHXMJXWY submitted 2025-08-28 cs.CV cs.AI

Zhizhong Huang , Xiaoming Liu This is my paper
classification cs.CVcs.AI
keywords generalizableobjectre-identificationin-contextlearningvisualprompttuningvisionfoundationmodelslargelanguageDINOv2few-shotidentitymatchingShopID10K
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

Object re-identification (ReID) today means training a separate model for each category — persons, vehicles, pets — and performance collapses when the model meets a category it never saw. VICP challenges this by claiming a model trained once on a few 'seen' categories can directly match instances of any unseen category, using only a few example pairs as prompts at deployment, with no parameter updates. The key move is to split the problem: a frozen large language model reasons over a handful of same-instance/different-instance pairs to infer what actually defines identity for that category (the logo, not the lighting), and this semantic judgment is translated into visual prompts injected into every layer of a frozen DINOv2 backbone, steering its attention toward identity-sensitive details. On unseen categories, the authors report that VICP beats both self-supervised baselines and fine-tuned few-shot baselines across six benchmarks, including a new e-commerce dataset, ShopID10K (10K instances, 34 categories), introduced to make cross-category generalization measurable at scale. The stated boundary: deployment assumes the category is already known, because comparing across categories can never retrieve the right match.

What carries the argument

The central object is the dynamic visual prompt: a small set of learnable tokens that the frozen LLM contextualizes against the in-context example pairs; a two-layer MLP then projects their hidden states into Ptask, the vision model's token dimension. At every layer of the frozen DINOv2 transformer these tokens are concatenated into the token sequence, so self-attention reweights spatial features toward identity-sensitive local details (logos, textures, stitching) and away from irrelevant variation (lighting, background, viewpoint). Supporting machinery: a Q-Former connector compresses each example image into 32 latent tokens so pairs fit the LLM context window; the in-context loss supervise

What would settle it

Shuffle the support-set labels (or replace the frozen LLM's context with random tokens) while keeping every other component identical, and measure mAP on unseen categories: if performance does not drop, the LLM's semantic inference is not what carries the gain. A complementary test: pick a novel category whose identity cues cannot be verbalized (two near-identical objects differing only by a tiny scratch) — the paper's account predicts VICP still beats the fine-tuned baseline there, and a tie would mark the boundary of the semantic channel.

Watch

Extended reading notes

Core claim

VICP claims to make object re-identification generalizable: a model trained on a few seen categories matches instances of any unseen category without updating any parameters. Given a support set of same-instance (positive) and different-instance (negative) pairs from the new category, a frozen LLM (LLaMA in the experiments) infers which visual attributes define identity — e.g., logo placement rather than lighting — and the contextualized tokens are projected into dynamic visual prompts injected into every transformer layer of a frozen DINOv2. Trained once on seen categories with a label-masked in-context loss, triplet loss, and an optimal-transport patch-alignment loss, the framework reports

Load-bearing premise

At deployment the category of every query and gallery image is assumed already known (the paper's 'category-aware inference'): if a detector supplies the wrong category, or the object's category is unknown, the framework can only compare within the assumed category, where the authors note a correct match is impossible — so 'generalizable' means known-category, within-category matching.

Editorial extensions

If this is right

  • Deploying ReID to a new object category reduces to collecting a few dozen labeled same-or-different pairs: the model generates category-specific visual prompts once, caches them, and reuses them for every query-gallery comparison — no retraining or per-dataset model needed.
  • A single frozen DINOv2 backbone, never updated for the target category, can outperform fine-tuned and few-shot baselines on unseen categories across benchmarks, including person and vehicle datasets the model was not trained for.
  • The design choices the paper credits for this — masking the LLM loss to label tokens, triplet loss for soft updates, and patch-level optimal-transport alignment — define a recipe for adapting foundation models to identity-level tasks without destroying their generalization.
  • ShopID10K, with 10K instances across 34 everyday categories under real e-commerce lighting, occlusion, and background variation, gives the community a testbed for cross-category ReID beyond the 180-instance CUTE lab dataset.

Reading between the lines

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

  • Because the method's prompts encode the LLM's verbalized rules, categories whose identity cues resist language — microscopic defects, worn textures, near-invisible seam differences — are the likely failure frontier; a targeted benchmark of such categories would map where the semantic channel stops helping.
  • The paper's category-aware assumption means real-world gains depend on the upstream detector being right: feeding mis-detected or out-of-vocabulary categories into the pipeline would test how gracefully the framework degrades, since the authors concede cross-category comparisons can never be correct.
  • The same prompt-conditioning interface could transfer to other instance-level tasks — speaker matching in audio, instance retrieval in 3D scans — wherever a frozen encoder can be steered by an LLM that verbalizes what differs between instances of the same kind.
  • A direct ablation the paper does not run would isolate the LLM's contribution: replace the LLM's context with random tokens or shuffled labels while keeping the visual-prompt machinery intact; if mAP holds, the gain is prompt tuning, not semantic reasoning.
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 / 6 minor

Summary. The paper introduces VICP, a framework for generalizable object re-identification. The model is trained on a set of base object categories and then, at test time, is given a small support set of positive/negative image pairs from a novel category. A frozen LLM processes visual tokens (compressed by a Q-Former) and label embeddings, together with learnable prompt tokens; the LLM's hidden states are projected by a visual head into visual prompts that are injected into every layer of a frozen DINOv2 ViT. The model is trained with a composite loss combining a triplet loss, an in-context learning loss on label tokens, and a patch-level optimal-transport alignment loss. The paper also introduces ShopID10K, a new e-commerce dataset of 10K instances across 34 categories, and evaluates the method on PetFace, MVImageNet, CUTE, MSMT17, Market1501, and VeRi-776. The central claim is that LLMs infer semantic identity-discriminative rules from the few-shot pairs and encode these rules into dynamic visual prompts, enabling parameter-free generalization to novel categories.

Significance. If the central mechanism were established, the paper would make a valuable contribution to a relatively underexplored problem: generalizable object ReID across arbitrary categories. The introduced ShopID10K dataset is a useful resource, addressing a gap in existing benchmarks that are either single-category or small-scale/lab-controlled. The paper also provides a systematic empirical comparison against many self-supervised and fine-tuned baselines, with consistent (if modest) improvements on the proposed task. The use of a frozen DINOv2 backbone and zero-parameter adaptation at test time is practically appealing. However, the paper's headline claim—that LLM-derived semantic rules are what drive the generalization—is not directly supported by the experiments. The reported gains of the full method over the strong Triplet+ baseline are 1.4–3.7 mAP across the main benchmarks, and the ablation isolating the LLM's contribution is inconclusive. The work would be strengthened by ablation of the LLM component itself, explicit values for unreported loss weights, and a better-specified protocol for the person/vehicle experiments.

major comments (4)
  1. The central claim that the LLM infers semantic identity rules and that these rules drive the visual-prompt gains is not verified. The paper never decodes the LLM's output, shows no example of an inferred rule, and includes no ablation replacing the frozen LLM with a non-LLM sequence model or a randomly initialized transformer. In Tab. 6, adding ICL visual prompts over Triplet+ yields only +1.0 mAP on PetFace and +1.5 mAP on ShopID10K, while direct LLM inference ([v]) is far below Triplet. This leaves open the possibility that the gains come from the trainable Q-Former, prompt tokens, and losses rather than from the LLM's semantic reasoning. A direct comparison (e.g., replacing the LLM with a frozen randomly initialized transformer, or a shallow non-pretrained sequence model) is needed to support the mechanism claimed in the abstract and introduction.
  2. The loss weights λICL and λalign are never reported, and no sensitivity analysis is provided. Since the composite loss in Eq. (7) determines the relative influence of the in-context objective and the patch-alignment objective, omitting these values makes the training procedure irreproducible. If, for example, λICL is effectively zero, the ICL loss would not train the connector/LLM pathway as described. Please report these values and their effect on the main results.
  3. The person/vehicle ReID experiments are underspecified. MSMT17, Market1501, and VeRi-776 contain a single category, so the 'novel category' setting does not directly apply. It is unclear what support set is used at test time, how the model was trained (on which base categories) for these benchmarks, and why Triplet+ (the strongest baseline elsewhere) is not included. Without this information, the comparison to PASS, TransReID, and other category-specific methods is not interpretable and does not clearly support the generalization claim.
  4. The hyperparameters K (number of in-context examples) and N (number of Q-Former tokens) are ablated on the novel-category test sets. The manuscript reports choosing K=64 and N=32 based on the best mAP on PetFace and ShopID10K, which are the same categories used to report final performance. This creates a form of test-set tuning bias; the generalization claim would be more credible if these hyperparameters were selected on a validation split or fixed a priori. At minimum, please clarify whether the reported numbers are the best over the shown hyperparameter grid or the actual held-out performance at the chosen values.
minor comments (6)
  1. The 'category-aware inference' assumption is clearly stated, and it is reasonable for a ReID pipeline with an upstream detector. However, the abstract's phrase 'arbitrary object categories' overstates the scope: the method requires knowing the category of every query and gallery image. Consider tempering the abstract phrasing or explicitly stating this limitation at the start.
  2. Fig. 1 shows 'Chimp' and 'Backpack' as unseen categories, but the task definition says categories are disjoint from base categories. This is fine, but the figure caption could clarify that the model never sees chimps or backpacks during training.
  3. The table heading says 'MVImageNe' (typo). Also, in the same section, the text says 'results on MVImageNe'.
  4. The text says 'We show thee ablation studies'—should be 'three'.
  5. 'ShopeID10K (ours)' should be 'ShopID10K (ours)' for consistency.
  6. For PetFace, the paper reports both verification (AUC/ACC) and identification (mAP/Top-1). It would be helpful to state explicitly that the identification protocol uses one image per query/gallery and how duplicate IDs across the train/test categories are handled, since PetFace originally has a different evaluation protocol.

Circularity Check

1 steps flagged · score 2.0 of 10

Mild test-set hyperparameter selection; core few-shot prompt generation is independent.

  1. fitted input called prediction [Sec. 4.3 Ablation Study (Table 6) and Sec. 4 Implementation Details]
    "Our method achieves peak performance at 64 examples, with 32 examples yielding marginally lower results and 128 examples causing degradation. ... The best performance is achieved at N = 32. ... During training, we randomly sample 64 positive/negative pairs. The number of visual tokens for Q-former is 32."

    The hyperparameters K (number of in-context pairs) and N (number of Q-Former latent tokens) are chosen by comparing mAP/Rank-1 on the same novel-category test sets (PetFace and ShopID10K) that are later reported as evidence of generalization. The reported advantage of VICP is therefore partially the result of selecting the configuration that maximizes performance on the evaluation set, rather than a prediction from a fixed, pre-specified model. This is a mild form of test-set circularity; it does not by itself make the whole approach equivalent to its inputs, since the prompt generator is trained only on base categories.

full rationale

The paper's central mechanism is a conditional prompt generator: a frozen LLM processes support-set pairs, and a trainable connector/visual head produces prompts injected into a frozen DINOv2. Training uses base-category pairs and losses (Eqs. 2, 5, 6); at test time only the support set from the novel category conditions the prompts. This is a legitimate few-shot generalization setup and is not self-definitional. The 'LLM infers semantic identity rules' narrative is not directly validated (no decoded rules, no random-transformer ablation), but that is a missing-mechanism/verification gap, not a circular derivation. The only concrete circularity is the selection of K=64 and N=32 on the novel-category test sets used for the final numbers; this is a mild tuning bias and warrants a score of 2 rather than 0. No load-bearing self-citations, imported uniqueness theorems, or ansatz-via-citation chains were found; the method is compared against external baselines and its gains are not forced by construction.

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

The paper introduces no new physical entities. Its free parameters are training hyperparameters, two of which (K, N) are tuned via ablation on the same novel-category test split. The most load-bearing assumptions are the category-aware inference premise, the noisy reviewer-as-ID labeling rule for ShopID10K, and the unverified claim that the LLM produces semantic identity rules.

free parameters (4)
  • loss weights λICL and λalign
    Composite loss in Sec. 3.4; the paper gives no values or sensitivity analysis for these balancing coefficients.
  • number of in-context examples K = 64
    K=64 selected via ablation on the novel category test sets (Tab. 6); 32 and 128 both perform worse.
  • number of latent tokens N (Q-Former) = 32
    N=32 selected via ablation; 16 and 64 degrade performance, indicating tuning on test data.
  • triplet margin α = 0.1
    Set in implementation details without justification or sensitivity study.
assumptions (4)
  • domain assumption At deployment, the object category is known a priori for every query and gallery image.
    Stated in Sec. 3.1 as a foundational assumption; without it, cross-category comparisons are undefined and the method cannot retrieve across categories.
  • ad hoc to paper Images uploaded by the same Amazon reviewer for a product share the same instance ID in ShopID10K.
    Dataset construction in Sec. 4; the benchmark's instance labels rest on this heuristic with no validation of label noise.
  • domain assumption The LLM can infer useful identity-discriminative rules from Q-Former-compressed visual tokens of positive/negative pairs.
    The method's explanatory claim; the paper provides no analysis of what these rules are, only indirect performance comparisons.
  • domain assumption DINOv2 features can be adaptively reweighted by injected visual prompts without destroying generalization.
    Relies on prior visual prompt tuning work; the paper shows empirical gains but offers no theoretical grounding.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalizable Object Re-Identification via Visual In-Context Prompting." pith.science (2026). https://pith.science/paper/UHXMJXWY

@misc{pith2026250821222,
  author       = {Pith},
  title        = {Pith review of: Generalizable Object Re-Identification via Visual In-Context Prompting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UHXMJXWY}},
  note         = {Machine review of arXiv:2508.21222}
}
read the original abstract

Current object re-identification (ReID) methods train domain-specific models (e.g., for persons or vehicles), which lack generalization and demand costly labeled data for new categories. While self-supervised learning reduces annotation needs by learning instance-wise invariance, it struggles to capture \textit{identity-sensitive} features critical for ReID. This paper proposes Visual In-Context Prompting~(VICP), a novel framework where models trained on seen categories can directly generalize to unseen novel categories using only \textit{in-context examples} as prompts, without requiring parameter adaptation. VICP synergizes LLMs and vision foundation models~(VFM): LLMs infer semantic identity rules from few-shot positive/negative pairs through task-specific prompting, which then guides a VFM (\eg, DINO) to extract ID-discriminative features via \textit{dynamic visual prompts}. By aligning LLM-derived semantic concepts with the VFM's pre-trained prior, VICP enables generalization to novel categories, eliminating the need for dataset-specific retraining. To support evaluation, we introduce ShopID10K, a dataset of 10K object instances from e-commerce platforms, featuring multi-view images and cross-domain testing. Experiments on ShopID10K and diverse ReID benchmarks demonstrate that VICP outperforms baselines by a clear margin on unseen categories. Code is available at https://github.com/Hzzone/VICP.

Figures

Figures reproduced from arXiv: 2508.21222 by the authors.

Figure 1
Figure 1. Overview of Object Re-Identification: (a) Traditional [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed framework. a) In-Context Visual Prompt Generation: Given few-shot input pairs, a frozen LLM [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Examples of 4 categories, 2 instances per category, and [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The ranking results of different models on PetFace/MVImageNet/ShopID10K/CUTE datasets. Blue/Red boxes indicate [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

79 extracted references · 57 canonical work pages

  1. [1]

    Flamingo: a visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in neural information processing systems, 2022

    work page 2022

  2. [2]

    Deep vit features as dense visual descriptors

    Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021

    arXiv 2021

  3. [3]

    Unicom: Uni- versal and compact representation learning for image retrieval

    Xiang An, Jiankang Deng, Kaicheng Yang, Jiawei Li, Ziyong Feng, Jia Guo, Jing Yang, and Tongliang Liu. Unicom: Uni- versal and compact representation learning for image retrieval. In ICLR, 2023

    work page 2023

  4. [4]

    Self-supervised learning from images with a joint-embedding predictive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bo- janowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619–15629, 2023

    work page 2023

  5. [5]

    Openflamingo: An open- source framework for training large autoregressive vision- language models

    Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An open- source framework for training large autoregressive vision- language models. arXiv preprint arXiv:2308.01390, 2023

    arXiv 2023

  6. [6]

    Rasa: Relation and sensitivity aware representation learning for text-based person search

    Yang Bai, Min Cao, Daming Gao, Ziqiang Cao, Chen Chen, Zhenfeng Fan, Liqiang Nie, and Min Zhang. Rasa: Relation and sensitivity aware representation learning for text-based person search. arXiv preprint arXiv:2305.13653, 2023

    arXiv 2023

  7. [7]

    Learning local feature descriptors with triplets and shallow convolutional neural networks

    Vassileios Balntas, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Learning local feature descriptors with triplets and shallow convolutional neural networks. In Bmvc, vol- ume 1, page 3, 2016

    work page 2016

  8. [8]

    Beit: Bert pre-training of image transformers

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021

    arXiv 2021

Show all 79 references
  1. [9]

    Unsupervised learning of visual features by contrasting cluster assignments

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020

    2020

  2. [10]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021

    2021

  3. [11]

    Wildlifedatasets: An open-source toolkit for ani- mal re-identification

    V ojtˇech ˇCerm´ak, Lukas Picek, Luk´aˇs Adam, and Kostas Pa- pafitsoros. Wildlifedatasets: An open-source toolkit for ani- mal re-identification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5953– 5963, 2024

    2024

  4. [12]

    Towards modality- agnostic person re-identification with descriptive query

    Cuiqun Chen, Mang Ye, and Ding Jiang. Towards modality- agnostic person re-identification with descriptive query. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15128–15137, 2023

    2023

  5. [13]

    Abd-net: Attentive but diverse person re-identification

    Tianlong Chen, Shaojin Ding, Jingyi Xie, Ye Yuan, Wuyang Chen, Yang Yang, Zhou Ren, and Zhangyang Wang. Abd-net: Attentive but diverse person re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8351–8361, 2019

    2019

  6. [14]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PmLR, 2020

    2020

  7. [15]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern re...

    2023

  8. [16]

    Im- proved baselines with momentum contrastive learning

    Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Im- proved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020

    2003 arXiv

  9. [17]

    Salience-guided cascaded suppression network for person re-identification

    Xuesong Chen, Canmiao Fu, Yong Zhao, Feng Zheng, Jingkuan Song, Rongrong Ji, and Yi Yang. Salience-guided cascaded suppression network for person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3300–3310, 2020

    2020

  10. [18]

    An empirical study of training self-supervised vision transformers

    Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9640–9649, 2021

    2021

  11. [19]

    Arcface: Additive angular margin loss for deep face recog- nition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recog- nition. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 4690–4699, 2019

    2019

  12. [20]

    A survey on in-context learning

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022

    2022 arXiv

  13. [21]

    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

  14. [22]

    Towards multimodal in-context learning for vision & language models

    Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid Karlinsky. Towards multimodal in-context learning for vision & language models. arXiv preprint arXiv:2403.12736, 2024

    2024 arXiv

  15. [23]

    Model- agnostic meta-learning for fast adaptation of deep networks

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model- agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–

  16. [24]

    Dreamsim: Learning new dimensions of human visual similarity using synthetic data

    Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023

    2023 arXiv

  17. [25]

    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

    2022

  18. [26]

    Language-guided hierarchical fine-grained image forgery de- tection and localization

    Xiao Guo, Xiaohong Liu, Iacopo Masi, and Xiaoming Liu. Language-guided hierarchical fine-grained image forgery de- tection and localization. In International Journal of Computer Vision, December 2024

    2024

  19. [27]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. CVPR, 2022. 9

    2022

  20. [28]

    Girshick

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual repre- sentation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726–9735, 2019

    2020

  21. [29]

    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

    2021

  22. [30]

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

    Weizhen He, Yiheng Deng, Shixiang Tang, Qihao Chen, Qingsong Xie, Yizhou Wang, Lei Bai, Feng Zhu, Rui Zhao, Wanli Ouyang, et al. Instruct-reid: A multi-purpose person re-identification task with instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat...

    2024

  23. [31]

    Clothing status awareness for long-term person re-identification

    Yan Huang, Qiang Wu, JingSong Xu, Yi Zhong, and ZhaoXi- ang Zhang. Clothing status awareness for long-term person re-identification. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 11895–11904, 2021

    2021

  24. [32]

    Learning representation for clustering via prototype scattering and positive sampling

    Zhizhong Huang, Jie Chen, Junping Zhang, and Hongming Shan. Learning representation for clustering via prototype scattering and positive sampling. IEEE Transactions on Pat- tern Analysis and Machine Intelligence , 45(6):7509–7524, 2022

    2022

  25. [33]

    Openclip, July

    Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Ha- jishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July

  26. [34]

    Visual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European conference on computer vision, pages 709–727. Springer, 2022

    2022

  27. [35]

    Dinov2 meets text: A unified framework for image-and pixel-level vision- language alignment

    Cijo Jose, Th´eo Moutakanni, Dahyun Kang, Federico Baldas- sarre, Timoth´ee Darcet, Hu Xu, Daniel Li, Marc Szafraniec, Micha¨el Ramamonjisoa, Maxime Oquab, et al. Dinov2 meets text: A unified framework for image-and pixel-level vision- language alignment. arXiv preprint arXiv:...

    2024 arXiv

  28. [36]

    Supervised contrastive learning

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661–18673, 2020

    2020

  29. [37]

    Adaface: Quality adaptive margin for face recognition

    Minchul Kim, Anil K Jain, and Xiaoming 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

    2022

  30. [38]

    Sapiensid: Foundation for human recognition

    Minchul Kim, Dingqiang Ye, Yiyang Su, Feng Liu, and Xi- aoming Liu. Sapiensid: Foundation for human recognition. In In Proceeding of IEEE Computer Vision and Pattern Recogni- tion, Nashville, TN, June 2025

    2025

  31. [39]

    Are these the same apple? comparing images based on object intrinsics

    Klemen Kotar, Stephen Tian, Hong-Xing Yu, Dan Yamins, and Jiajun Wu. Are these the same apple? comparing images based on object intrinsics. In Advances in Neural Information Processing Systems, 2023

    2023

  32. [40]

    Klemen Kotar, Stephen Tian, Hong-Xing Yu, Daniel L. K. Yamins, and Jiajun Wu. Are These the Same Apple? Com- paring Images Based on Object Intrinsics, Nov. 2023

    2023

  33. [41]

    Blip- 2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip- 2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  34. [42]

    Farsight: A physics-driven whole-body biometric sys- tem at large distance and altitude

    Feng Liu, Ryan Ashbaugh, Nicholas Chimitt, Najmul Has- san, Ali Hassani, Ajay Jaiswal, Minchul Kim, Zhiyuan Mao, Christopher Perry, Zhiyuan Ren, Yiyang Su, Pegah Varghaei, Kai Wang, Xingguang Zhang, Stanley Chan, Arun Ross, Humphrey Shi, Zhangyang Wang, Anil Jain, and Xiaoming...

    2024

  35. [43]

    Learning clothing and pose invariant 3d shape representa- tion for long-term person re-identification

    Feng Liu, Minchul Kim, ZiAng Gu, Anil Jain, and Xiaoming Liu. Learning clothing and pose invariant 3d shape representa- tion for long-term person re-identification. InIn Proceeding of International Conference on Computer Vision, Paris, France, October 2023

    2023

  36. [44]

    Distilling clip with dual guidance for learning discriminative human body shape representation

    Feng Liu, Minchul Kim, Zhiyuan Ren, and Xiaoming Liu. Distilling clip with dual guidance for learning discriminative human body shape representation. In In Proceeding of IEEE Computer Vision and Pattern Recognition, Seattle, W A, June 2024

    2024

  37. [45]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024

    2024

  38. [46]

    Learning memory-augmented unidirectional metrics for cross-modality person re-identification

    Jialun Liu, Yifan Sun, Feng Zhu, Hongbin Pei, Yi Yang, and Wenhui Li. Learning memory-augmented unidirectional metrics for cross-modality person re-identification. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19366–19375, 2022

    2022

  39. [47]

    Multiple instance learning via iterative self-paced supervised contrastive learning

    Kangning Liu, Weicheng Zhu, Yiqiu Shen, Sheng Liu, Narges Razavian, Krzysztof J Geras, and Carlos Fernandez-Granda. Multiple instance learning via iterative self-paced supervised contrastive learning. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern R...

    2023

  40. [48]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023

    2023 arXiv

  41. [49]

    A deep learning-based approach to progressive vehicle re- identification for urban surveillance

    Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. A deep learning-based approach to progressive vehicle re- identification for urban surveillance. In Computer Vision– ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pag...

    2016

  42. [50]

    Bag of tricks and a strong baseline for deep person re-identification

    Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019

    2019

  43. [51]

    Boult, Anderson Rocha, Haidong Zhu, Zhaoheng 10 Zheng, Ram Nevatia, Zaigham Randhawa, Sinan Sabri, and Gianfranco Doretto

    Kien Nguyen, Clinton Fookes, Sridha Sridharan, Feng Liu, Xiaoming Liu, Arun Ross, Dana Michalski, Huy Nguyen, Debayan Deb, Mahak Kothari, Manisha Saini, Dawei Du, Scott McCloskey, Gabriel Bertocco, Fernanda Andal´o, Ter- rance E. Boult, Anderson Rocha, Haidong Zhu, Zhaoheng 10...

    2023

  44. [52]

    Deep metric learning via lifted structured feature embedding

    Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on com- puter vision and pattern recognition, pages 4004–4012, 2016

    2016

  45. [53]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth´ee Darcet, Th´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  46. [54]

    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

  47. [55]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Doll ´ar, and Christoph Feicht...

    2024 arXiv

  48. [56]

    Contrastive learning with hard negative sam- ples

    Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Ste- fanie Jegelka. Contrastive learning with hard negative sam- ples. arXiv preprint arXiv:2010.04592, 2020

    2010 arXiv

  49. [57]

    Petface: A large-scale dataset and benchmark for animal identification

    Risa Shinoda and Kaede Shiohara. Petface: A large-scale dataset and benchmark for animal identification. In European Conference on Computer Vision, pages 19–36. Springer, 2024

    2024

  50. [58]

    Prototypi- cal networks for few-shot learning

    Jake Snell, Kevin Swersky, and Richard Zemel. Prototypi- cal networks for few-shot learning. In Advances in neural information processing systems, 2017

    2017

  51. [59]

    Hamobe: Hierarchical and adaptive mixture of biometric experts for video-based person reid

    Yiyang Su, Yunping Shi, Feng Liu, and Xiaoming Liu. Hamobe: Hierarchical and adaptive mixture of biometric experts for video-based person reid. In In Proceeding of Inter- national Conference on Computer Vision, Honolulu, Hawaii, October 2025

    2025

  52. [60]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Mar- tinet, Marie-Anne Lachaux, Timoth ´ee Lacroix, Baptiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation lan...

    2023 arXiv

  53. [61]

    Learning discriminative features with multiple granularities for person re-identification

    Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In Proceedings of the 26th ACM international conference on Multimedia, pages 274–282, 2018

    2018

  54. [62]

    Panet: Few-shot image semantic segmentation with prototype alignment

    Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision , pages 9197– 9206, 2019

    2019

  55. [63]

    Contrastive learning with stronger augmentations

    Xiao Wang and Guo-Jun Qi. Contrastive learning with stronger augmentations. IEEE transactions on pattern analy- sis and machine intelligence, 45(5):5549–5560, 2022

    2022

  56. [65]

    Person transfer gan to bridge domain gap for person re-identification

    Longhui Wei, Shiliang Zhang, Wen Gao, and Qi 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

    2018

  57. [66]

    Dreamrelation: Relation-centric video customization

    Yujie Wei, Shiwei Zhang, Hangjie Yuan, Biao Gong, Longxi- ang Tang, Xiang Wang, Haonan Qiu, Hengjia Li, Shuai Tan, Yingya Zhang, et al. Dreamrelation: Relation-centric video customization. arXiv preprint arXiv:2503.07602, 2025

    2025 arXiv

  58. [67]

    Unsupervised feature learning via non-parametric instance discrimination

    Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733–3742, 2018

    2018

  59. [68]

    A fast proximal point method for computing exact wasserstein distance

    Yujia Xie, Xiangfeng Wang, Ruijia Wang, and Hongyuan Zha. A fast proximal point method for computing exact wasserstein distance. In Uncertainty in artificial intelligence, pages 433–

  60. [69]

    Simmim: A simple framework for masked image modeling

    Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653–9663, 2022

    2022

  61. [70]

    Learning with twin noisy labels for visible-infrared person re-identification

    Mouxing Yang, Zhenyu Huang, Peng Hu, Taihao Li, Jiancheng Lv, and Xi Peng. Learning with twin noisy labels for visible-infrared person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14308–14317, 2022

    2022

  62. [71]

    Biggait: Learning gait representation you want by large vision models

    Dingqiang Ye, Chao Fan, Jingzhe Ma, Xiaoming Liu, and Shiqi Yu. Biggait: Learning gait representation you want by large vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 200–210, 2024

    2024

  63. [72]

    Mvimgnet: A large- scale dataset of multi-view images

    Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large- scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recog- ...

    2023

  64. [73]

    Vehicle re-identification for automatic video traffic surveillance

    Dominik Zapletal and Adam Herout. Vehicle re-identification for automatic video traffic surveillance. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 25–31, 2016

    2016

  65. [74]

    A survey of large language models

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023

    2023 arXiv

  66. [75]

    Scalable person re-identification: A benchmark

    Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jing- dong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international confer- ence on computer vision, pages 1116–1124, 2015

    2015

  67. [76]

    Pass: Part-aware self-supervised pre- training for person re-identification

    Kuan Zhu, Haiyun Guo, Tianyi Yan, Yousong Zhu, Jinqiao Wang, and Ming Tang. Pass: Part-aware self-supervised pre- training for person re-identification. In European conference on computer vision, pages 198–214. Springer, 2022. 11

    2022

  68. [77]

    Aaformer: Auto-aligned transformer for person re-identification

    Kuan Zhu, Haiyun Guo, Shiliang Zhang, Yaowei Wang, Jing Liu, Jinqiao Wang, and Ming Tang. Aaformer: Auto-aligned transformer for person re-identification. IEEE Transactions on Neural Networks and Learning Systems, 2023

    2023

  69. [78]

    VL-ICL bench: The devil in the details of multimodal in- context learning

    Yongshuo Zong, Ondrej Bohdal, and Timothy Hospedales. VL-ICL bench: The devil in the details of multimodal in- context learning. In The Thirteenth International Conference on Learning Representations, 2025. 12 Generalizable Object Re-Identification via Visual In-Context Prompt...

    2025

  70. [80]

    While ShopID10K is visually observed with occlusion variations, there are no explicit occlusion labels, making it difficult to quantitatively evaluate occlusion robustness

    Additional Experiments Robustness to occlusions: Robustness to pose/lighting has been validated through MVImageNet and CUTE datasets, which offer rich pose variations via multi-view videos or lab-controlled pose/lighting variations. While ShopID10K is visually observed with oc...

  71. [2021]

    If you use this software, please cite it as below

Pith tools

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