Pith. sign in

REVIEW 4 major objections 5 minor 86 references

Person search reaches a new state of the art using a frozen diffusion model as the backbone.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 12:46 UTC pith:DKYLY3AR

load-bearing objection DiffPS is a plausible first use of frozen diffusion features for person search, with strong ablations, but SFAN's cosine similarity as written is dimensionally undefined and the SOTA margins rest on test-set-tuned hyperparameters without repeated runs. the 4 major comments →

arxiv 2510.01841 v2 pith:DKYLY3AR submitted 2025-10-02 cs.CV

Leveraging Prior Knowledge of Diffusion Model for Person Search

Giyeol Kim , Sooyoung Yang , Jihyong Oh , Myungjoo Kang , Chanho Eom This is my paper
classification cs.CV
keywords person searchdiffusion modelsperson re-identificationobject detectionfrozen backbonesemantic attentionfrequency refinementgenerative priors
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that a frozen text-to-image diffusion model can serve as the visual backbone for person search—locating and identifying people in uncropped scenes—and that doing so dissolves the long-standing trade-off between detection and re-identification. The method, DiffPS, keeps a Stable Diffusion UNet entirely frozen, so detection and re-ID branches no longer pull shared weights in opposite directions. Three lightweight modules extract what the backbone has to offer: attention maps for localization, wavelet-based frequency refinement to counter the model's shape bias, and text-aligned semantic part features to enrich identity descriptions. On the two standard benchmarks, CUHK-SYSU and PRW, DiffPS reports the best published numbers, including a large jump in mAP over prior ImageNet-based systems. The authors' central thesis is that generative pre-training encodes the spatial context and fine-grained identity cues that person search needs, and that these can be harvested without fine-tuning.

Core claim

The central claim is that the internal representations of a pre-trained text-to-image diffusion model, kept completely frozen, are superior to ImageNet-trained backbones for both sub-tasks of person search, and that freezing the backbone removes the optimization conflict between them. Concretely, DiffPS uses the UNet of a latent diffusion model at timestep t=0 and extracts features from selected up-stage layers. Four observed properties of these features motivate the design: cross-attention maps aligned with text tokens such as 'person' provide localization cues; feature informativeness varies strongly with timestep, peaking at t=0; up-stage layers retain both fine detail and global context;

What carries the argument

The central object is the frozen UNet of a latent text-to-image diffusion model (Stable Diffusion v2-1), probed at timestep t=0 and at carefully chosen up-stage layers. It does two kinds of work at once: as a feature extractor it supplies spatial detail for detection and fine-grained appearance for re-ID; as a fixed parameter set it mechanically removes gradient interference, since neither task updates the backbone. Around it, three modules carry the method: DGRPN thresholds the 'person' cross-attention map and models peaks as Gaussians to modulate detection features; MSFRN applies discrete wavelet decomposition with sub-band attention to re-balance high frequencies undone by the model's sha

Load-bearing premise

The whole approach rests on the premise that a frozen text-to-image UNet, read at timestep zero, delivers features that retain fine-grained identity cues alongside spatial layout; if those features are too generative-specific, too noisy, or too biased toward low frequencies to be salvaged by the three readout modules, the reported gains disappear.

What would settle it

Run DiffPS with the diffusion UNet replaced by a randomly initialized (or a plain VAE) encoder of the same architecture, keeping all modules and training identical; if the performance gap to the reported numbers is small, the diffusion pre-training is not the cause. Alternatively, fix the hyperparameters selected on PRW and evaluate on a disjoint person-search benchmark; large drops would indicate that the choices overfit the test set.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Because the backbone is frozen, detection and re-ID can be trained separately or jointly with identical results, eliminating the need for the gradient-balancing heuristics used by previous one-step person search methods.
  • If the reported numbers hold, a generative pre-trained model is a viable alternative to ImageNet or self-supervised backbones for tasks that need both localization and instance-level discrimination, not just generation.
  • The layer/timestep selection rules (up-stage levels 2–3, t=0) provide a recipe for anyone reusing diffusion features in other recognition tasks.
  • The semantic part-attention mechanism (SFAN) improves robustness on occluded person queries, per the paper's occlusion experiments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Looking past this paper, the same frozen-backbone pattern could be applied to other joint detection-and-recognition problems, but the practical benefit will hinge on the inference cost of a full UNet forward pass versus the accuracy gain; a distilled or pruned generative model would be the natural stress test.
  • The paper's own ablations show that the bare frozen backbone (row (e), Table 11) underperforms the final system by a wide margin, suggesting the diffusion prior is necessary but not sufficient—the modules do much of the heavy lifting, so the claim that 'diffusion priors solve person search' should be read as 'diffusion priors plus task-specific readout modules'.
  • Because the authors pick hyperparameters (feature combination, τ, text prompts) on the PRW test set, a stronger test of the approach would be to freeze all hyperparameters and evaluate on a held-out scene dataset; the paper does not report such an evaluation.
  • The choice of t=0 means the UNet is used essentially as a fixed encoder rather than a denoiser; it would be informative to compare against the VAE encoder alone to see how much of the gain comes from the UNet's learned structure versus the latent space itself.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes DiffPS, a person-search framework that replaces an ImageNet-pretrained backbone with a frozen Stable Diffusion v2-1 UNet, and introduces three modules: DGRPN (cross-attention-guided region proposal), MSFRN (multi-scale frequency refinement for high-frequency detail), and SFAN (text-aligned semantic feature aggregation). The authors claim that freezing the backbone eliminates the detection/re-ID optimization conflict, and report state-of-the-art results on CUHK-SYSU (97.8 mAP / 98.4 Top-1) and PRW (62.0 mAP / 91.0 Top-1). The paper includes extensive ablations, a backbone-controlled comparison (Table 11), layer-wise and timestep analyses, and an occlusion robustness study.

Significance. If the reported results hold, the paper makes a useful contribution in two respects: it demonstrates that a frozen generative backbone can serve both detection and re-ID in person search, and it decouples the two sub-tasks by construction, offering a clean solution to an optimization problem that prior decoupled methods (e.g., DMRNet) address only partially. The manuscript is strengthened by careful ablation tables (Tables 4, 5, 11), a controlled comparison against the same SD backbone for existing methods, and falsifiable analyses such as the timestep and layer-wise studies. These go beyond a simple 'diffusion features are good' claim. However, as detailed below, the current write-up leaves several load-bearing implementation details unspecified, one formula is dimensionally undefined as written, and there is a direct inconsistency in the reported hyperparameter value for τ. These issues must be resolved before the claimed SOTA can be independently verified.

major comments (4)
  1. [Sec. 3.6 (SFAN)] The cosine similarity between the refined feature map F_refined and CLIP text embeddings is not realizable as written. F_refined is a UNet up-stage feature with channel dimension 320/640/1280 for SD v2-1, while the CLIP text embeddings used in SD v2-1 have dimension 1024 (77×1024 tensor). No projection layer is described in the text or in Fig. 4(d), so 'the cosine similarity between the refined feature map and each text embedding token' is dimensionally undefined unless C=1024, which is not the case for the levels used. Without a learned or fixed projection to a common space, the spatial semantic maps S_c cannot be computed, and the +0.5 mAP / +0.3 Top-1 gains in Table 5 cannot be attributed to the described mechanism. Please specify the projection (e.g., 1×1 conv or MLP mapping F_refined to 1024-dim) and its training, or revise the description to match the actual implementation.
  2. [Sec. 4.1 (Implementation details)] The input pipeline for the Stable Diffusion backbone is not specified. The paper says 'We use Stable Diffusion v2-1 as the backbone' but does not state whether raw images are VAE-encoded into latent space, at what resolution, whether the UNet is run with a single forward pass at timestep t=0 (as implied by Sec. 4.3 and Fig. 5), or which cross-attention layer provides M_att for DGRPN (Sec. 3.4 says only 'the corresponding cross-attention layer'). These details are essential for reimplementation. Please specify them explicitly, including the text-conditioning setup (e.g., empty prompt or 'person' prompt) for the backbone forward pass.
  3. [Sec. 4.1 vs. Sec. 4.3 / Table 4] There is a direct inconsistency in the DGRPN threshold: Sec. 4.1 states 'the threshold τ is set to 0.7', but Sec. 4.3 and Table 4 report that 'The optimal performance is achieved at τ=0.5' and the final configuration's detection numbers in Table 11 (AP 94.8 / Recall 98.1) match the τ=0.5 row, not the τ=0.7 row (94.7 / 97.6). Please correct the stated implementation value or the table; this inconsistency prevents knowing which configuration produced the reported SOTA.
  4. [Tables 3, 4, 7 and Fig. 5] The final configuration is selected using PRW test-set metrics, with no validation split or repeated runs. The margins for the selection decisions are small: 0.7 mAP between (g) and (f) in Table 3, 0.1 AP / 0.5 Recall between τ=0.5 and τ=0.8 in Table 4, and 0.3–0.5 mAP in Table 7. Without a validation-based selection or variance estimates, the reported SOTA gains may partly reflect test-set overfitting. Please report selection on a validation split (or on CUHK-SYSU) and/or repeated runs (seeds) for the key configurations.
minor comments (5)
  1. [Fig. 4 caption] The caption contains typographical artifacts ('SemanIc-adapIve Feature AggregaIon Network') and Sec. 4.3 has a grammar error ('a widely used in prior methods'). Please proofread.
  2. [Table 3 caption] The caption says '(g) combines 1, 2, and 4' but does not specify that these are the numbers of features taken from levels 1, 2, and 3 respectively, which would help the reader map the row to Sec. 3.5's description of seven feature maps.
  3. [Fig. 5] The figure does not identify which curves correspond to which layers/feature combinations beyond the caption. Add a legend or state in the text which of the three layers (blue/orange/green in (a)) and which feature sets ((g), (b), (c) from Table 3 in (b)) are used.
  4. [Sec. 3.6] Prompts such as 'head' and 'shirts' may tokenize into multiple tokens in CLIP. Please specify whether S_c is computed from a single token, a mean-pooled embedding, or the full prompt embedding, and how exactly the four maps are obtained.
  5. [General] The paper does not state whether source code will be released. Given the number of unspecified implementation details (VAE encoding, cross-attention layer selection, projection in SFAN), a public implementation would substantially strengthen reproducibility.

Circularity Check

0 steps flagged

No circular derivation: DiffPS uses an external frozen Stable Diffusion backbone, and its modules are evaluated on external benchmarks. Mild test-set selection and reproducibility gaps are methodological risks, not circularity.

full rationale

DiffPS's claimed derivation chain is not circular. The backbone (SD v2-1 UNet) is an external pretrained model, and the three modules (DGRPN, MSFRN, SFAN) are trained with standard detection and re-ID losses and ablated on external benchmarks (CUHK-SYSU and PRW); no equation defines a module output in terms of the headline metric, and no fitted parameter is renamed as a prediction. The claimed perfect decoupling (Table 2) is a structural consequence of freezing the backbone, but it is still empirically verified and is not a derived prediction that equals its own input. The only author self-citation ([14], Eom & Ham, for stripe-based partitioning) is incidental and not load-bearing; there is no self-citation chain, imported uniqueness theorem, or ansatz smuggled in via citation. I do flag two non-circular concerns: (1) several configuration choices are selected on the PRW test set — feature combination (g) in Table 3, text prompts in Table 7, DGRPN threshold in Table 4, and timestep t=0 in Fig. 5 — and the same PRW test numbers are then reported as the headline result (Sec. 4.2); this is test-set selection bias rather than circular derivation; (2) Sec. 3.6 specifies cosine similarity between UNet feature maps F_refined and CLIP text embeddings without stating a projection to a common dimension, which is a reproducibility gap, not a circularity. The Limitation section (Suppl. K) only acknowledges computational overhead and does not admit a circular step. Neither issue makes the derivation equivalent to its inputs, so the circularity score stays in the 0-2 band.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper's central empirical gains rest on five tunable hyperparameters selected on the PRW test set, an under-specified input pipeline for the frozen Stable Diffusion UNet, and the domain assumption that generative features transfer to person search without fine-tuning.

free parameters (5)
  • DGRPN threshold tau = 0.7 (implementation) / 0.5 (ablation optimum)
    Threshold on cross-attention map for Gaussian centers; chosen via PRW ablation in Table 4.
  • DGRPN minimum Gaussian std delta = 5
    Ablated in Suppl. I; controls minimum spatial extent of Gaussian proposals.
  • SFAN text prompts = "head", "shirts", "pants", "shoes"
    Selected by test-set ablation in Table 7; clothing terms chosen because they gave highest mAP.
  • Re-ID feature-map selection = 4 maps from level 3 + 2 from level 2 + 1 from level 1 (combination g)
    Selected on PRW test set via Table 3; used in final configuration.
  • Diffusion timestep t = 0
    Figure 5 and Sec. 4.3 report optimal features at t=0 on PRW test; used for all final results.
axioms (4)
  • domain assumption Stable Diffusion v2-1 UNet features, without task-specific fine-tuning, are a strong backbone for person detection and re-ID.
    The central premise; supported only by empirical comparison (Table 9 vs DINO and Table 1 vs ImageNet backbones).
  • domain assumption The pre-trained UNet receives person-search images and produces the described feature maps F_u^l via the standard Stable Diffusion pipeline (including VAE encoding and t=0).
    Not detailed in the paper; Sec. 4.1 omits input preprocessing, yet all experiments depend on it.
  • ad hoc to paper Choosing architecture and hyperparameters on the PRW test set is a valid way to report final SOTA.
    Tables 3, 4, and 7 use PRW test metrics to select feature sets, tau, and text prompts; final Table 1 uses the same test set.
  • domain assumption DWT/IDWT sub-band refinement enhances identity-relevant high-frequency features enough to justify MSFRN.
    MSFRN design and shape-bias validation (Table 12, Fig. 8) assume frequency decomposition improves re-ID.

pith-pipeline@v1.3.0-alltime-deepseek · 24520 in / 11667 out tokens · 90931 ms · 2026-08-04T12:46:20.677530+00:00 · methodology

0 comments
read the original abstract

Person search aims to jointly perform person detection and re-identification by localizing and identifying a query person within a gallery of uncropped scene images. Existing methods predominantly utilize ImageNet pre-trained backbones, which may be suboptimal for capturing the complex spatial context and fine-grained identity cues necessary for person search. Moreover, they rely on a shared backbone feature for both person detection and re-identification, leading to suboptimal features due to conflicting optimization objectives. In this paper, we propose DiffPS (Diffusion Prior Knowledge for Person Search), a novel framework that leverages a pre-trained diffusion model while eliminating the optimization conflict between two sub-tasks. We analyze key properties of diffusion priors and propose three specialized modules: (i) Diffusion-Guided Region Proposal Network (DGRPN) for enhanced person localization, (ii) Multi-Scale Frequency Refinement Network (MSFRN) to mitigate shape bias, and (iii) Semantic-Adaptive Feature Aggregation Network (SFAN) to leverage text-aligned diffusion features. DiffPS sets a new state-of-the-art on CUHK-SYSU and PRW.

Figures

Figures reproduced from arXiv: 2510.01841 by Chanho Eom, Giyeol Kim, Jihyong Oh, Myungjoo Kang, Sooyoung Yang.

Figure 1
Figure 1. Figure 1: (a) PCA [44] visualization of feature maps, comparing widely-adopted ImageNet [10] pre-trained backbones (ResNet50 [20] and ConvNeXt [41]) in person search with a diffusion model. The diffusion features exhibit richer spatial context and fine-grained details. (b) Existing methods [3, 6, 18, 24, 27, 34, 74, 77, 78] share backbone features for detection and re-ID, causing optimization conflicts due to compet… view at source ↗
Figure 2
Figure 2. Figure 2: Detailed architecture of the UNet [58] in diffusion mod￾els [55, 57], comprising down-stage, mid-stage, and up-stage for hierarchical feature processing. (Best viewed in color.) sis of diffusion priors for person search (Sec. 3.2). Finally, we present our proposed framework (Sec. 3.3). 3.1. Preliminaries Diffusion models. Diffusion models [23, 55, 57, 62] gen￾erate images by iteratively refining Gaussian n… view at source ↗
Figure 3
Figure 3. Figure 3: (a) Cross-attention maps highlighting different semantic [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Overview of DiffPS framework. DiffPS leverages a pre-trained diffusion model’s UNet as the backbone, with three specialized [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Timestep analysis of diffusion features on PRW [ [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results of MSFRN and SFAN. (a) The [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative and quantitative comparison of small person detection performance. Left: Quantitative comparison of AP [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Cue-conflict examples with shape/texture labels and [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Feature map visualization from Res [20] modules across different stages and levels of UNet [58]. The visualizations are generated using PCA [44] on the output feature maps, with each row showing a different level and each column representing different res modules within that level. The input image is shown at the top for reference. Colors indicate the intensity and pattern of feature activations, demon￾str… view at source ↗
Figure 10
Figure 10. Figure 10: Visualization of feature characteristics across different timesteps in the diffusion process. Visualization using PCA [ [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

86 extracted references · 12 linked inside Pith

  1. [1]

    The eurocity persons dataset: A novel benchmark for object detection.arXiv preprint arXiv:1805.07193, 2018

    Markus Braun, Sebastian Krebs, Fabian Flohr, and Dariu M Gavrila. The eurocity persons dataset: A novel benchmark for object detection.arXiv preprint arXiv:1805.07193, 2018. 3

    Pith/arXiv arXiv 2018

  2. [2]

    From handcrafted to deep features for pedestrian detection: A survey.IEEE transactions on pattern analysis and machine intelligence, 44(9):4913–4934, 2021

    Jiale Cao, Yanwei Pang, Jin Xie, Fahad Shahbaz Khan, and Ling Shao. From handcrafted to deep features for pedestrian detection: A survey.IEEE transactions on pattern analysis and machine intelligence, 44(9):4913–4934, 2021. 1

    2021

  3. [3]

    Pstr: End-to-end one-step person search with trans- formers

    Jiale Cao, Yanwei Pang, Rao Muhammad Anwer, Hisham Cholakkal, Jin Xie, Mubarak Shah, and Fahad Shahbaz Khan. Pstr: End-to-end one-step person search with trans- formers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9458– 9467, 2022. 1, 2, 3, 6, 7

    2022

  4. [4]

    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. InPro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 2

    2021

  5. [5]

    Person search via a mask-guided two-stream cnn model

    Di Chen, Shanshan Zhang, Wanli Ouyang, Jian Yang, and Ying Tai. Person search via a mask-guided two-stream cnn model. InProceedings of the european conference on com- puter vision (ECCV), pages 734–750, 2018. 1, 2

    2018

  6. [6]

    Norm-aware embedding for efficient person search

    Di Chen, Shanshan Zhang, Jian Yang, and Bernt Schiele. Norm-aware embedding for efficient person search. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12615–12624, 2020. 1, 2, 3, 6, 7

    2020

  7. [7]

    Your diffusion model is secretly a certifiably robust classifier.arXiv preprint arXiv:2402.02316, 2024

    Huanran Chen, Yinpeng Dong, Shitong Shao, Zhongkai Hao, Xiao Yang, Hang Su, and Jun Zhu. Your diffusion model is secretly a certifiably robust classifier.arXiv preprint arXiv:2402.02316, 2024. 2, 3, 4

    Pith/arXiv arXiv 2024

  8. [8]

    Dif- fusiondet: Diffusion model for object detection

    Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Dif- fusiondet: Diffusion model for object detection. InProceed- ings of the IEEE/CVF international conference on computer vision, pages 19830–19843, 2023. 3

    2023

  9. [9]

    Beyond appear- ance: a semantic controllable self-supervised learning frame- work for human-centric visual tasks

    Weihua Chen, Xianzhe Xu, Jian Jia, Hao Luo, Yaohua Wang, Fan Wang, Rong Jin, and Xiuyu Sun. Beyond appear- ance: a semantic controllable self-supervised learning frame- work for human-centric visual tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15050–15061, 2023. 3, 7

    2023

  10. [10]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 1, 2, 4

    2009

  11. [11]

    Bi-directional interaction network for person search

    Wenkai Dong, Zhaoxiang Zhang, Chunfeng Song, and Tie- niu Tan. Bi-directional interaction network for person search. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2839–2848, 2020. 7

    2020

  12. [12]

    Instance guided proposal network for person search

    Wenkai Dong, Zhaoxiang Zhang, Chunfeng Song, and Tie- niu Tan. Instance guided proposal network for person search. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2585–2594, 2020. 1, 2

    2020

  13. [13]

    An image is worth 16x16 words: Transformers for image recognition at scale.ICLR, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale.ICLR, 2021. 4, 1, 2, 5, 7

    2021

  14. [14]

    Learning disentangled rep- resentation for robust person re-identification.Advances in neural information processing systems, 32, 2019

    Chanho Eom and Bumsub Ham. Learning disentangled rep- resentation for robust person re-identification.Advances in neural information processing systems, 32, 2019. 5

    2019

  15. [15]

    Unsuper- vised pre-training for person re-identification

    Dengpan Fu, Dongdong Chen, Jianmin Bao, Hao Yang, Lu Yuan, Lei Zhang, Houqiang Li, and Dong Chen. Unsuper- vised pre-training for person re-identification. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14750–14759, 2021. 3

    2021

  16. [16]

    Wichmann, and Wieland Brendel

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, 2022. 3

    2022

  17. [17]

    Re-id driven lo- calization refinement for person search

    Chuchu Han, Jiacheng Ye, Yunshan Zhong, Xin Tan, Chi Zhang, Changxin Gao, and Nong Sang. Re-id driven lo- calization refinement for person search. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 9814–9823, 2019. 1, 2

    2019

  18. [18]

    Decoupled and memory-reinforced net- works: Towards effective feature learning for one-step per- son search

    Chuchu Han, Zhedong Zheng, Changxin Gao, Nong Sang, and Yi Yang. Decoupled and memory-reinforced net- works: Towards effective feature learning for one-step per- son search. InProceedings of the AAAI Conference on Ar- tificial Intelligence, pages 1505–1512, 2021. 1, 2, 3, 5, 6, 7

    2021

  19. [19]

    Chuchu Han, Zhedong Zheng, Kai Su, Dongdong Yu, Ze- huan Yuan, Changxin Gao, Nong Sang, and Yi Yang. Dmr- net++: Learning discriminative features with decoupled net- works and enriched pairs for one-step person search.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7319–7337, 2023. 7

    2023

  20. [20]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 1, 2, 4, 7, 5, 6

    2016

  21. [21]

    Rethinking im- agenet pre-training

    Kaiming He, Ross Girshick, and Piotr Doll´ar. Rethinking im- agenet pre-training. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 4918–4927,

  22. [22]

    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. InProceedings of the IEEE/CVF international conference on computer vision, pages 15013–15022, 2021. 1

    2021

  23. [23]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 3

    2020

  24. [24]

    Gallery filter network for person search

    Lucas Jaffe and Avideh Zakhor. Gallery filter network for person search. InProceedings of the IEEE/CVF Winter Con- ference on Applications of Computer Vision, pages 1684– 1693, 2023. 1, 2, 3, 5, 7, 8

    2023

  25. [25]

    Intrigu- ing properties of generative classifiers.arXiv preprint arXiv:2309.16779, 2023

    Priyank Jaini, Kevin Clark, and Robert Geirhos. Intrigu- ing properties of generative classifiers.arXiv preprint arXiv:2309.16779, 2023. 4

    Pith/arXiv arXiv 2023

  26. [26]

    Psdiff: Diffusion model for person search with iterative and collaborative refinement,

    Chengyou Jia, Minnan Luo, Zhuohang Dang, Guang Dai, Xiaojun Chang, and Jingdong Wang. Psdiff: Diffusion model for person search with iterative and collaborative refinement,

  27. [27]

    Scene-adaptive person search via bilateral mod- ulations.arXiv preprint arXiv:2405.02834, 2024

    Yimin Jiang, Huibing Wang, Jinjia Peng, Xianping Fu, and Yang Wang. Scene-adaptive person search via bilateral mod- ulations.arXiv preprint arXiv:2405.02834, 2024. 1, 2, 3, 5, 7, 8

    Pith/arXiv arXiv 2024

  28. [28]

    Prototype-guided saliency feature learning for per- son search

    Hanjae Kim, Sunghun Joung, Ig-Jae Kim, and Kwanghoon Sohn. Prototype-guided saliency feature learning for per- son search. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4865– 4874, 2021. 7

    2021

  29. [29]

    Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma. Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980, 2014. 7

    Pith/arXiv arXiv 2014

  30. [30]

    Text- to-image diffusion models are great sketch-photo matchmak- ers

    Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Text- to-image diffusion models are great sketch-photo matchmak- ers. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 16826–16837,

  31. [31]

    Oimnet++: Prototypical normaliza- tion and localization-aware learning for person search

    Sanghoon Lee, Youngmin Oh, Donghyeon Baek, Junghyup Lee, and Bumsub Ham. Oimnet++: Prototypical normaliza- tion and localization-aware learning for person search. In European Conference on Computer Vision, pages 621–637. Springer, 2022. 5

    2022

  32. [32]

    Your diffusion model is secretly a zero-shot classifier

    Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 2206–2217,

  33. [33]

    Deep- reid: Deep filter pairing neural network for person re- identification

    Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang. Deep- reid: Deep filter pairing neural network for person re- identification. In2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 152–159, 2014. 3

    2014

  34. [34]

    Sequential end-to-end net- work for efficient person search

    Zhengjia Li and Duoqian Miao. Sequential end-to-end net- work for efficient person search. InProceedings of the AAAI Conference on Artificial Intelligence, pages 2011– 2019, 2021. 1, 2, 3, 5, 6, 7, 8

    2011

  35. [35]

    Open-vocabulary object segmenta- tion with diffusion models

    Ziyi Li, Qinye Zhou, Xiaoyun Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Open-vocabulary object segmenta- tion with diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 7667– 7676, 2023. 2, 3, 4

    2023

  36. [36]

    Per- son re-identification by local maximal occurrence represen- tation and metric learning

    Shengcai Liao, Yang Hu, Xiangyu Zhu, and Stan Z Li. Per- son re-identification by local maximal occurrence represen- tation and metric learning. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 2197–2206, 2015. 1

    2015

  37. [37]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740–755. Springer, 2014. 4

    2014

  38. [38]

    Focal loss for dense object detection, 2018

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll´ar. Focal loss for dense object detection, 2018. 1

    2018

  39. [39]

    High-level semantic feature detection: A new perspective for pedestrian detection

    Wei Liu, Shengcai Liao, Weiqiang Ren, Weidong Hu, and Yinan Yu. High-level semantic feature detection: A new perspective for pedestrian detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5187–5196, 2019. 1

    2019

  40. [40]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 7

    2021

  41. [41]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 11976–11986,

  42. [42]

    Diffusion hyperfeatures: Searching through time and space for semantic correspondence.Ad- vances in Neural Information Processing Systems, 36, 2024

    Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holyn- ski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence.Ad- vances in Neural Information Processing Systems, 36, 2024. 2, 3, 4

    2024

  43. [43]

    Diffusionseg: Adapting diffusion towards unsupervised object discovery

    Chaofan Ma, Yuhuan Yang, Chen Ju, Fei Zhang, Jinxiang Liu, Yu Wang, Ya Zhang, and Yanfeng Wang. Diffusionseg: Adapting diffusion towards unsupervised object discovery. arXiv preprint arXiv:2303.09813, 2023. 2, 3

    Pith/arXiv arXiv 2023

  44. [44]

    Principal components analysis (pca).Computers & Geosciences, 19 (3):303–342, 1993

    Andrzej Ma ´ckiewicz and Waldemar Ratajczak. Principal components analysis (pca).Computers & Geosciences, 19 (3):303–342, 1993. 1, 4, 8, 6, 7

    1993

  45. [45]

    Not all diffusion model activa- tions have been evaluated as discriminative features.arXiv preprint arXiv:2410.03558, 2024

    Benyuan Meng, Qianqian Xu, Zitai Wang, Xiaochun Cao, and Qingming Huang. Not all diffusion model activa- tions have been evaluated as discriminative features.arXiv preprint arXiv:2410.03558, 2024. 2, 3, 4

    Pith/arXiv arXiv 2024

  46. [46]

    Do text-free diffusion models learn discriminative vi- sual representations? InEuropean Conference on Computer Vision, pages 253–272

    Soumik Mukhopadhyay, Matthew Gwilliam, Yosuke Yam- aguchi, Vatsal Agarwal, Namitha Padmanabhan, Archana Swaminathan, Tianyi Zhou, Jun Ohya, and Abhinav Shrivas- tava. Do text-free diffusion models learn discriminative vi- sual representations? InEuropean Conference on Computer Vision, pages 253–272. Springer, 2024. 4

    2024

  47. [47]

    Query-guided end-to-end person search

    Bharti Munjal, Sikandar Amin, Federico Tombari, and Fabio Galasso. Query-guided end-to-end person search. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2019. 5, 7

    2019

  48. [48]

    Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation.Advances in Neural Information Processing Systems, 36, 2024

    Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation.Advances in Neural Information Processing Systems, 36, 2024. 2

    2024

  49. [49]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021. 2, 3

    Pith/arXiv arXiv 2021

  50. [50]

    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. 2

    Pith/arXiv arXiv 2023

  51. [51]

    Mask- guided attention network for occluded pedestrian detection

    Yanwei Pang, Jin Xie, Muhammad Haris Khan, Rao Muham- mad Anwer, Fahad Shahbaz Khan, and Ling Shao. Mask- guided attention network for occluded pedestrian detection. InProceedings of the IEEE/CVF international conference on computer vision, pages 4967–4975, 2019. 1

    2019

  52. [52]

    Ld-znet: A latent diffusion ap- proach for text-based image segmentation

    Koutilya Pnvr, Bharat Singh, Pallabi Ghosh, Behjat Sid- diquie, and David Jacobs. Ld-znet: A latent diffusion ap- proach for text-based image segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 4157–4168, 2023. 2, 3, 4

    2023

  53. [53]

    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. InInternational conference on machine learning, pages 8748–8763. PMLR, 2021. 3, 6

    2021

  54. [54]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. InInternational confer- ence on machine learning, pages 8821–8831. Pmlr, 2021. 2, 3, 4

    2021

  55. [55]

    Hierarchical text-conditional image gener- ation with clip latents.arXiv preprint arXiv:2204.06125, 1 (2):3, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with clip latents.arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 3, 4, 6

    Pith/arXiv arXiv 2022

  56. [56]

    Faster r-cnn: Towards real-time object detection with region proposal networks.Advances in neural information process- ing systems, 28, 2015

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks.Advances in neural information process- ing systems, 28, 2015. 5, 8, 1

    2015

  57. [57]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2, 3, 4, 6, 7

    2022

  58. [58]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. InMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234–241. Springer, 2015. 3, 4, 1, 6, 7

    2015

  59. [59]

    Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022. 3, 4

    2022

  60. [60]

    Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in Neural In- formation Processing Systems, 35:25278–25294, 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Worts- man, et al. Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in Neural In- formation Processing Systems, 35:25278–25294, 2022. 2, 3

    2022

  61. [61]

    Crowdhuman: A bench- mark for detecting human in a crowd.arXiv preprint arXiv:1805.00123, 2018

    Shuai Shao, Zijian Zhao, Boxun Li, Tete Xiao, Gang Yu, Xiangyu Zhang, and Jian Sun. Crowdhuman: A bench- mark for detecting human in a crowd.arXiv preprint arXiv:1805.00123, 2018. 3

    Pith/arXiv arXiv 2018

  62. [62]

    De- noising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. De- noising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 3

    Pith/arXiv arXiv 2010

  63. [63]

    Pose-driven deep convolutional model for person re-identification

    Chi Su, Jianing Li, Shiliang Zhang, Junliang Xing, Wen Gao, and Qi Tian. Pose-driven deep convolutional model for person re-identification. InProceedings of the IEEE inter- national conference on computer vision, pages 3960–3969,

  64. [64]

    Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline), 2018

    Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline), 2018. 1

    2018

  65. [65]

    Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline)

    Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). InProceed- ings of the European conference on computer vision (ECCV), pages 480–496, 2018. 5

    2018

  66. [66]

    Diffuse attend and segment: Un- supervised zero-shot segmentation using stable diffusion

    Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse attend and segment: Un- supervised zero-shot segmentation using stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 3554–3563, 2024. 2, 3

    2024

  67. [67]

    Divide and conquer: Hybrid pre-training for person search

    Yanling Tian, Di Chen, Yunan Liu, Jian Yang, and Shanshan Zhang. Divide and conquer: Hybrid pre-training for person search. InProceedings of the AAAI Conference on Artificial Intelligence, pages 5224–5232, 2024. 2, 3

    2024

  68. [68]

    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. InProceedings of the 26th ACM international conference on Multimedia, page 274–282. ACM, 2018. 1, 5

    2018

  69. [69]

    Domain generalization via frequency-domain-based feature disentanglement and inter- action

    Jingye Wang, Ruoyi Du, Dongliang Chang, Kongming Liang, and Zhanyu Ma. Domain generalization via frequency-domain-based feature disentanglement and inter- action. InProceedings of the 30th ACM international con- ference on multimedia, pages 4821–4829, 2022. 4

    2022

  70. [70]

    Object pose estimation via the aggregation of diffusion features

    Tianfu Wang, Guosheng Hu, and Hongguang Wang. Object pose estimation via the aggregation of diffusion features. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 10238–10247, 2024. 2

    2024

  71. [71]

    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. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 79–88, 2018. 3

    2018

  72. [72]

    Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using dif- fusion models

    Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using dif- fusion models. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 1206–1217,

  73. [73]

    Ian: the individual aggre- gation network for person search.Pattern Recognition, 87: 332–340, 2019

    Jimin Xiao, Yanchun Xie, Tammam Tillo, Kaizhu Huang, Yunchao Wei, and Jiashi Feng. Ian: the individual aggre- gation network for person search.Pattern Recognition, 87: 332–340, 2019. 7

    2019

  74. [74]

    Joint detection and identification feature learn- ing for person search

    Tong Xiao, Shuang Li, Bochao Wang, Liang Lin, and Xiao- gang Wang. Joint detection and identification feature learn- ing for person search. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3415– 3424, 2017. 1, 2, 3, 5, 6, 7, 8

    2017

  75. [75]

    Open-vocabulary panop- tic segmentation with text-to-image diffusion models

    Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiao- long Wang, and Shalini De Mello. Open-vocabulary panop- tic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 2955–2966, 2023. 2, 3, 4

    2023

  76. [76]

    Denoiserep: Denoising model for representation learn- ing, 2024

    Zhengrui Xu, Guan’an Wang, Xiaowen Huang, and Jitao Sang. Denoiserep: Denoising model for representation learn- ing, 2024. 3

    2024

  77. [77]

    Anchor-free person search

    Yichao Yan, Jinpeng Li, Jie Qin, Song Bai, Shengcai Liao, Li Liu, Fan Zhu, and Ling Shao. Anchor-free person search. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7690–7699, 2021. 1, 2, 3, 6, 7

    2021

  78. [78]

    Cas- cade transformers for end-to-end person search

    Rui Yu, Dawei Du, Rodney LaLonde, Daniel Davila, Christopher Funk, Anthony Hoogs, and Brian Clipp. Cas- cade transformers for end-to-end person search. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7267–7276, 2022. 1, 2, 3, 5, 6, 7

    2022

  79. [79]

    Pha: Patch-wise high-frequency augmentation for transformer-based person re-identification

    Guiwei Zhang, Yongfei Zhang, Tianyu Zhang, Bo Li1, and Shiliang Pu. Pha: Patch-wise high-frequency augmentation for transformer-based person re-identification. InProceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 14133–14142, 2023. 4

    2023

  80. [80]

    A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence.Advances in Neural Information Processing Systems, 36, 2024

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Pola- nia Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence.Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4

    2024

Showing first 80 references.