Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Expanding zero-shot pedestrian attribute training sets with image-to-image diffusion-generated images improves mean accuracy on three benchmarks, with the largest gain on RAPzs (75.23 vs 71.98).

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 ↗

Diffusion-based synthetic expansion of pedestrian attribute training sets yields 1.1 to 3.3 point mA gains over the base PAR model on PA100k, PETAzs, and RAPzs.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A useful empirical sweep of diffusion-based PAR augmentation, but the central claim is undercut by an uncontrolled doubling of training set size. the 3 major comments →

arxiv 2509.02161 v1 pith:JBC2EIDC submitted 2025-09-02 cs.CV

Enhancing Zero-Shot Pedestrian Attribute Recognition with Synthetic Data Generation: A Comparative Study with Image-To-Image Diffusion Models

Pablo Ayuso-Albizu , Juan C. SanMiguel , Pablo Carballeira This is my paper
classification cs.CV
keywords pedestrian attribute recognitionzero-shot pedestrian identitysynthetic data augmentationimage-to-image diffusionlatent diffusion modelstextual inversionFIDdata expansion
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.

The reading

Pedestrian attribute recognition models need many annotated images, and zero-shot evaluation makes the problem harder by removing training identities from the test set. This paper tries to show that synthetic images produced by an image-to-image diffusion model can expand the training set and improve performance: the authors identify which prompt and image properties matter, compare three recent diffusion-augmentation techniques, and pick the best one to double the size of three zero-shot benchmarks. Their central result is that the expanded dataset beats both the base model and six standard augmentation methods on all three datasets, with the largest gain on RAPzs (75.23 vs 71.98 mean accuracy). The paper also reports that prompt alignment is critical and that blurry or context-poor conditioning images degrade generation quality. If the claim holds, it gives PAR practitioners a practical recipe for using diffusion models to reduce the cost of manual annotation and to improve generalization to unseen identities.

Core claim

The paper's discovery is a recipe-level finding: for pedestrian attribute recognition under zero-shot identity evaluation, expanding the training set with image-to-image diffusion images labeled from their prompts improves accuracy beyond what standard augmentations achieve. Using a per-attribute textual inversion variant of DAFUSION with a hand-built prompt template that aligns with each dataset's attribute vocabulary, the authors generate one synthetic image per training image for PA100k, PETAzs, and RAPzs. The resulting mean accuracies are 81.58, 73.07, and 75.23, compared with 80.44, 71.47, and 71.98 for the base model, and the expansions beat Autoaug, CutMix, Mixup, RandAug, TrivialAugm

What carries the argument

The central mechanism is image-to-image latent diffusion conditioned on aligned attribute prompts, with per-attribute textual inversion. Textual inversion learns a small word-like token vector for an attribute so the diffusion model can generate images containing that attribute; the paper adapts DAFUSION so each attribute class gets its own token. A real pedestrian image is encoded to a latent, a prompt assembled from the image's annotated attributes guides denoising, and the resulting synthetic image is assigned the prompt's attributes as labels. The paper identifies four controls: prompt type and alignment, input image clarity and context, denoising strength (how much of the original image

Load-bearing premise

Synthetic training samples are labeled with the attributes written in their prompts, and those prompted attributes are actually present in the generated image; if image-to-image diffusion drops or changes an attribute, the expansion adds noisy labels and the measured gains may come from that noise rather than from useful synthetic data.

What would settle it

Train a reference PAR model on real data only, run it on the generated samples, and compare its attribute predictions with the prompt-derived labels; then retrain the pipeline keeping only synthetic samples whose predicted attributes match the prompt. If the match rate is low, or if filtering on it removes the reported mA improvements (75.23 vs 71.98 on RAPzs, 73.07 vs 71.47 on PETAzs, 81.58 vs 80.44 on PA100k), the central claim fails.

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

If this is right

  • On all three zero-shot benchmarks, the diffusion-expanded training set beats the base model and six standard augmentations (Autoaug, CutMix, Mixup, RandAug, TrivialAugment, Augmix) under the same architecture.
  • The largest per-attribute gains appear on underrepresented attributes such as handtruck, backpack, box, skirt, hat, calling, pushing, and carrying-by-arm, suggesting synthetic expansion mainly helps the tail of the attribute distribution.
  • Prompt alignment is a first-order factor: aligned dataset captions produce lower FID than handcrafted or LLM prompts in every strength/scale configuration tested.
  • Input image quality matters: increased blur and reduced spatial context raise FID, so applying the recipe to very low-resolution surveillance crops without preprocessing will likely weaken the gains.

Where Pith is reading between the lines

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

  • Because each synthetic sample is labeled only with attributes inside its prompt, the method cannot create new attribute combinations beyond the original annotation vocabulary; adding an external attribute proposer, such as a vision-language model, is a natural untested extension.
  • The post-generation FIDs (33-44) are still far above the train-test FIDs (7-13), so the performance gain appears to come from added diversity rather than photorealism; if so, even lighter-weight or less realistic generators that diversify more cheaply may help.
  • The paper trains one classifier architecture with one backbone; the same expansion may transfer to other PAR models or to neighboring tasks like person re-identification, but the paper does not test this.
  • Since DAFUSION tokens are learned per attribute from subsets of images containing that attribute, rare attributes have fewer conditioning images; balancing prompt sampling by attribute frequency is a testable way to push the reported rare-attribute gains further.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 5 minor

Summary. The paper studies the use of image-to-image diffusion models to generate synthetic pedestrian images for zero-shot Pedestrian Attribute Recognition (PAR). It proposes a conceptual decomposition of the diffusion pipeline into prompts, image properties, and configuration parameters (strength/scale), and tests these on FID-based image-quality experiments. Three recent diffusion-augmentation techniques—DAFUSION, DALDA, and GIF—are compared on RAPzs, with DAFUSION selected for dataset expansion. The final experiments train a PAR model on real data expanded with DAFUSION-generated synthetic images and report mean-accuracy (mA) improvements on PA100k, PETAzs, and RAPzs (Table 5) relative to the base method and several classical augmentation baselines.

Significance. If the reported gains are attributable to the synthetic data itself, the paper would provide a useful empirical result for PAR, a task where annotated data are scarce and where diffusion-based augmentation has been less explored than in generic classification. The study has strengths: it covers multiple prompt designs, image perturbations, and three diffusion frameworks; it uses standard zero-shot PAR benchmarks; and it reports per-attribute results in Figure 6. However, the central comparison in Table 5 is confounded by training-set size, and the paper does not provide code or checkpoints. The contribution is therefore currently more of a promising empirical exploration than a validated method.

major comments (3)
  1. [§5.5.1 and Table 5] The main claim that DAFUSION-based expansion improves zero-shot PAR is not supported by a controlled comparison. Section 5.5.1 states that the expansion 'doubl[es] the training set size,' while the augmentation baselines in Table 5 (Autoaug, CutMix, Mixup, RandAug, Trivial, Augmix) do not add new training samples. If the number of epochs is held fixed, the expanded model receives twice as many gradient updates and sample presentations, so the mA gains could reflect data quantity rather than synthetic-image quality. The paper must add a same-size control: for example, train the base method for twice as many epochs, or train it on the real data re-sampled to double the number of updates, and compare. Without this control, the reported 1–4% improvements do not uniquely demonstrate a benefit of the DAFUSION-generated images.
  2. [§5.5.1 and Figure 6] Synthetic labels are assigned directly from the prompt attributes without verifying that the generated image actually contains those attributes. The text says 'Each synthetic sample is labeled only with attributes contained in the associated prompt,' and prompts are built from the original image annotations. Since img2img diffusion can change appearance, the expansion may inject systematic label noise. This is especially relevant to the per-attribute improvements in Figure 6: attributes such as handtruck, backpack, and shoulderbag are rare in the real data, and the expansion may simply oversample them, whether or not the generated images are correctly labeled. I request a validation step: run the trained PAR model (or a separate attribute classifier) on the synthetic set to estimate label accuracy, or provide human evaluation of a sample of synthetic images. This would also help interpre
  3. [§5.5.1] The selection of the HiSt-HiSc configuration for the final expansion is not fully justified. In Table 4, DAFUSION achieves its best FID with LoSt-LoSc (135), while HiSt-HiSc gives 164. The paper argues that 'LoStr-LoSc setup yields minimal variation between inputs,' but no quantitative diversity or augmentation-benefit metric is provided. If FID is the stated quality criterion, the choice of a worse-FID configuration appears arbitrary; if diversity is the goal, the paper should measure it directly. A simple remedy is to run the final PAR experiment with more than one configuration, or to report a diversity metric (e.g., self-FID or LPIPS distance to the conditioning image) for each configuration.
minor comments (5)
  1. [§5.2 and Table 1] The term 'aligned' for prompts is not defined precisely. The text says alignment 'ensures that prompted attributes appear in the generated images,' but the verification method is not described. Please clarify whether this is done manually, by a classifier, or by construction of the prompt from existing captions.
  2. [§5.1 and Table 1] Typographical inconsistency: the configuration names appear as 'HiStr HiSc', 'HiStr LoSc', and 'LoStr LoSc' in Table 1, while the text and other tables use 'HiSt'/'LoSt'. Please standardize.
  3. [§4.1] The 'c g a 0' split of the MALS dataset is mentioned without a definition or reference. Please specify what this split denotes and why it was chosen.
  4. [§5.5.1] The reported FID values compare generated images to the training set, while the reference values are 'train-test FIDs.' The sentence is ambiguous about which reference is used for each number. Please state explicitly the distributions being compared.
  5. [Supplementary D.2] The training setup lists optimizer, learning-rate schedule, and loss, but does not state the number of epochs or the total number of iterations. This is important for assessing the training-set-size confound described above.

Circularity Check

0 steps flagged

No significant circularity: empirical augmentation study with external FID selection and real-test-set evaluation.

full rationale

The paper is an empirical study of diffusion-based data expansion for pedestrian attribute recognition. There is no formal derivation chain whose conclusions reduce to its inputs. The only potentially self-referential aspect is that synthetic images are labeled with attributes taken from the prompt, which in turn is built from the original image annotations. However, this is standard practice in dataset augmentation and does not make the reported mA gains circular: the target metric (mA) is measured on fixed real test sets and is never used to define or fit any component of the pipeline. The selection of DAFUSION over other techniques is based on FID, an external image-quality metric, not on the PAR metric. No uniqueness theorem is imported, no self-citation is load-bearing, and no fitted parameter is renamed as a prediction. The absence of a same-size training control is a possible confound (the expanded dataset doubles the number of training samples), but that is a validity limitation, not a circularity. The paper is self-contained against external benchmarks and its central claim is empirical rather than definitional. Therefore no circular step is identified.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The paper's results rest on two assumptions: that synthetic images inherit the attributes described in prompts, and that FID-based technique selection transfers to PAR accuracy. Neither is independently verified; the free parameters are the hand-picked diffusion configuration and prompt texts.

free parameters (2)
  • diffusion strength/scale (HiSt-HiSc) = 0.6 / 15
    Chosen by hand for dataset expansion based on qualitative arguments, despite LoSt-LoSc having better FID in Table 4.
  • hand-crafted prompt templates per dataset
    Separate templates are built for RAPzs, PETAzs, and PA100k (Supplementary A.1), and these exact wordings affect generation quality.
axioms (3)
  • domain assumption Synthetic images labeled by the attributes in the prompt actually contain those attributes
    Section 5.5.1 assigns each synthetic sample labels from the prompt; if generation fails to preserve attributes, the expansion injects label noise.
  • domain assumption FID between generated and real images is a reliable proxy for downstream PAR training utility
    Section 5.4 selects DAFUSION based on FID; the validity of this choice for the final mA metric is not tested.
  • standard math Standard Stable Diffusion v1.4 and the PAR training pipeline of Jia et al. [15] behave as reported
    The whole study rests on these publicly available models and datasets.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Zero-Shot Pedestrian Attribute Recognition with Synthetic Data Generation: A Comparative Study with Image-To-Image Diffusion Models." pith.science (2026). https://pith.science/paper/JBC2EIDC

@misc{pith2026250902161,
  author       = {Pith},
  title        = {Pith review of: Enhancing Zero-Shot Pedestrian Attribute Recognition with Synthetic Data Generation: A Comparative Study with Image-To-Image Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JBC2EIDC}},
  note         = {Machine review of arXiv:2509.02161}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Pedestrian Attribute Recognition (PAR) involves identifying various human attributes from images with applications in intelligent monitoring systems. The scarcity of large-scale annotated datasets hinders the generalization of PAR models, specially in complex scenarios involving occlusions, varying poses, and diverse environments. Recent advances in diffusion models have shown promise for generating diverse and realistic synthetic images, allowing to expand the size and variability of training data. However, the potential of diffusion-based data expansion for generating PAR-like images remains underexplored. Such expansion may enhance the robustness and adaptability of PAR models in real-world scenarios. This paper investigates the effectiveness of diffusion models in generating synthetic pedestrian images tailored to PAR tasks. We identify key parameters of img2img diffusion-based data expansion; including text prompts, image properties, and the latest enhancements in diffusion-based data augmentation, and examine their impact on the quality of generated images for PAR. Furthermore, we employ the best-performing expansion approach to generate synthetic images for training PAR models, by enriching the zero-shot datasets. Experimental results show that prompt alignment and image properties are critical factors in image generation, with optimal selection leading to a 4.5% improvement in PAR recognition performance.

Figures

Figures reproduced from arXiv: 2509.02161 by Juan C. SanMiguel, Pablo Ayuso-Albizu, Pablo Carballeira.

Figure 1
Figure 1. Figure 1: Proposed conceptual framework to study image-guided diffusion models for synthetic data augmentation in Pedestrian Attribute [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visual comparison between Text-to-Image and Image [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Examples of conditional images for blurring experi [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Examples of conditional images showing the pedes [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Examples of generated images. First row: init img from [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparative results in terms of mA for each attribute of [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visual examples of resolution experiments. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Example of aspect ratio experiments. C.2. Image to image based: context based In this subsection we wanted to show visual results from the context experiment presented in the main document. In the [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visual results of blurring experiments. Supplementary References [37] Jia et al., “Rethinking of pedestrian attribute recognition: A reliable evaluation under zero-shot pedestrian identity setting. arxiv 2021,” arXiv.2107.03576, vol. 2107. [38] Park et al., “Integrating language-derived appearance elements with visual cues in pedestrian detection,” IEEE Transactions on Circuits and Systems for Video Techno… view at source ↗
Figure 10
Figure 10. Figure 10: Visual results of context experiments [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visual examples of dataset expansion with each dataset explored. For RAPzs [15], PETAzs [15] and PA100k [14], first row [PITH_FULL_IMAGE:figures/full_fig_p012_11.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

39 extracted references · 31 canonical work pages · 1 internal anchor

  1. [1]

    Human attribute recognition—a compre- hensive survey,

    Yaghoubi et al., “Human attribute recognition—a compre- hensive survey,” Applied Sciences, vol. 10, no. 16, p. 5608, 2020

    work page 2020

  2. [2]

    Generative adversarial models for people at- tribute recognition in surveillance,

    Fabbri et al., “Generative adversarial models for people at- tribute recognition in surveillance,” inIEEE Int. Conf. on ad- vanced video and signal based surveillance, pp. 1–6, 2017

    work page 2017

  3. [3]

    Advances in diffusion models for image data augmentation: A review of methods, models, evaluation metrics and future research directions,

    Alimisis et al. , “Advances in diffusion models for image data augmentation: A review of methods, models, evaluation metrics and future research directions,”Artificial Intelligence Review, vol. 58, no. 4, pp. 1–55, 2025

    work page 2025

  4. [4]

    Is synthetic data from generative models ready for image recognition?,

    He et al., “Is synthetic data from generative models ready for image recognition?,” arXiv:2210.07574, 2022

    Pith/arXiv arXiv 2022

  5. [5]

    Diversify your vision datasets with automatic diffusion-based augmentation,

    Dunlap et al., “Diversify your vision datasets with automatic diffusion-based augmentation,” Advances in Neural Infor- mation Processing Systems, vol. 36, 2024

    work page 2024

  6. [6]

    Expanding small-scale datasets with guided imagination,

    Zhang et al., “Expanding small-scale datasets with guided imagination,” Advances in neural information processing systems, vol. 36, pp. 76558–76618, 2023

    work page 2023

  7. [9]

    High-resolution image synthesis with latent diffusion models,

    Rombach et al., “High-resolution image synthesis with latent diffusion models,” in IEEE/CVF Conf. on computer vision and pattern recognition, pp. 10684–10695, 2022

    work page 2022

  8. [10]

    Pedestrian attribute recognition: A survey,

    Wang et al. , “Pedestrian attribute recognition: A survey,” Pattern Recognition, vol. 121, p. 108220, 2022

    work page 2022

  9. [11]

    Pedestrian attribute recognition at far distance,

    Deng et al., “Pedestrian attribute recognition at far distance,” in Proc. of ACM Int. Conf. on Multimedia , pp. 789–792, 2014

    work page 2014

  10. [12]

    A richly annotated dataset for pedestrian attribute recognition,

    Li et al., “A richly annotated dataset for pedestrian attribute recognition,” arXiv:1603.07054, 2016

    Pith/arXiv arXiv 2016

  11. [13]

    A richly annotated pedestrian dataset for person re- trieval in real surveillance scenarios,

    Li et al., “A richly annotated pedestrian dataset for person re- trieval in real surveillance scenarios,” IEEE trans. on image processing, vol. 28, no. 4, pp. 1575–1590, 2018

    work page 2018

  12. [14]

    Hydraplus-net: Attentive deep features for pedes- trian analysis,

    Liu et al., “Hydraplus-net: Attentive deep features for pedes- trian analysis,” in IEEE Int. Conf. on computer vision , pp. 350–359, 2017

    work page 2017

  13. [16]

    Generative adversarial networks,

    Goodfellow et al., “Generative adversarial networks,” Com- munications of the ACM, vol. 63, no. 11, pp. 139–144, 2020

    work page 2020

  14. [17]

    A systematic review of generative adversarial net- works for traffic state prediction: overview, taxonomy, and future prospects,

    Li et al., “A systematic review of generative adversarial net- works for traffic state prediction: overview, taxonomy, and future prospects,” Information Fusion, p. 102915, 2025

    work page 2025

  15. [18]

    Diffusion models beat gans on image syn- thesis,

    Dhariwal et al., “Diffusion models beat gans on image syn- thesis,” Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021

    work page 2021

  16. [19]

    Sine: Single image editing with text-to-image diffusion models,

    Zhang et al., “Sine: Single image editing with text-to-image diffusion models,” in IEEE/CVF Conf. on Computer Vision and Pattern Recognition, pp. 6027–6037, 2023

    work page 2023

  17. [20]

    Cyclenet: Rethinking cycle consistency in text- guided diffusion for image manipulation,

    Xu et al., “Cyclenet: Rethinking cycle consistency in text- guided diffusion for image manipulation,” Advances in Neu- ral Information Processing Systems , vol. 36, pp. 10359– 10384, 2023

    work page 2023

  18. [21]

    Instance-aware data augmentation for unsuper- vised person re-identification,

    Dou et al., “Instance-aware data augmentation for unsuper- vised person re-identification,” in Int. Conf. on Information Tech. in Medicine and Education, pp. 425–429, IEEE, 2023

    work page 2023

  19. [22]

    An image is worth one word: Person- alizing text-to-image generation using textual inversion,

    Gal et al. , “An image is worth one word: Person- alizing text-to-image generation using textual inversion,” arXiv:2208.01618, 2022

    Pith/arXiv arXiv 2022

  20. [23]

    Synthesizing efficient data with diffusion models for person re-identification pre-training,

    Niu et al., “Synthesizing efficient data with diffusion models for person re-identification pre-training,”Machine Learning, vol. 114, no. 3, pp. 1–25, 2025

    work page 2025

  21. [24]

    Pose-dive: Pose-diversified augmenta- tion with diffusion model for person re-identification,

    Kim et al. , “Pose-dive: Pose-diversified augmenta- tion with diffusion model for person re-identification,” arXiv:2406.16042, 2024

    Pith/arXiv arXiv 2024

  22. [25]

    U-net: Convolutional networks for biomedical image segmentation,

    Ronneberger et al. , “U-net: Convolutional networks for biomedical image segmentation,” in MICCAI, Proc. Part III, 18th Int. Conf., pp. 234–241, Springer, 2015

    work page 2015

  23. [26]

    Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark,

    Yang et al., “Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark,” in 31st ACM Int. Conf. on Multimedia, pp. 4492–4501, 2023

    work page 2023

  24. [27]

    Tju-dhd: A diverse high-resolution dataset for object detection,

    Pang et al. , “Tju-dhd: A diverse high-resolution dataset for object detection,” IEEE Trans. Image Process., vol. 30, pp. 207–219, 2020

    work page 2020

  25. [28]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    Heusel et al., “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” Advances in neural information processing systems, vol. 30, 2017

    work page 2017

  26. [29]

    Integrating language-derived appearance ele- ments with visual cues in pedestrian detection,

    Park et al. , “Integrating language-derived appearance ele- ments with visual cues in pedestrian detection,”IEEE Trans- actions on Circuits and Systems for Video Technology, 2024

    work page 2024

  27. [30]

    Autoaugment: Learning augmentation strate- gies from data,

    Cubuk et al., “Autoaugment: Learning augmentation strate- gies from data,” in IEEE/CVF Conf. on computer vision and pattern recognition, pp. 113–123, 2019

    work page 2019

  28. [31]

    Cutmix: Regularization strategy to train strong classifiers with localizable features,

    Yun et al., “Cutmix: Regularization strategy to train strong classifiers with localizable features,” in IEEE/CVF Conf. on computer vision, pp. 6023–6032, 2019

    work page 2019

  29. [32]

    mixup: Beyond empirical risk minimization,

    Zhang et al., “mixup: Beyond empirical risk minimization,” arXiv:1710.09412, 2017

    Pith/arXiv arXiv 2017

  30. [33]

    Randaugment: Practical automated data aug- mentation with a reduced search space,

    Cubuk et al., “Randaugment: Practical automated data aug- mentation with a reduced search space,” in IEEE/CVF Int. Conf. on computer vision and pattern recognition work- shops, pp. 702–703, 2020

    work page 2020

  31. [34]

    Trivialaugment: Tuning-free yet state-of-the- art data augmentation,

    M ¨uller et al., “Trivialaugment: Tuning-free yet state-of-the- art data augmentation,” in IEEE/CVF Int. Conf. on computer vision, pp. 774–782, 2021

    work page 2021

  32. [35]

    Augmix: A simple data processing method to improve robustness and uncertainty,

    Hendrycks et al. , “Augmix: A simple data processing method to improve robustness and uncertainty,” arXiv preprint arXiv:1912.02781, 2019

    Pith/arXiv arXiv 1912

  33. [36]

    Deep residual learning for image recognition,

    He et al. , “Deep residual learning for image recognition,” in IEEE Conf. on computer vision and pattern recognition , pp. 770–778, 2016. SUPPLEMENTARY MATERIAL This supplementary material is provided to support and enhance the findings presented in the main manuscript. It includes additional figures, tables and additional experiments that could not be ac...

    work page 2016

  34. [37]

    Rethinking of pedestrian attribute recognition: A reliable evaluation under zero-shot pedestrian identity setting. arxiv 2021,

    Jia et al., “Rethinking of pedestrian attribute recognition: A reliable evaluation under zero-shot pedestrian identity setting. arxiv 2021,” arXiv.2107.03576, vol. 2107

    Pith/arXiv arXiv 2021

  35. [38]

    Integrating language-derived appearance elements with visual cues in pedestrian detection,

    Park et al. , “Integrating language-derived appearance elements with visual cues in pedestrian detection,” IEEE Transactions on Circuits and Systems for Video Technology, 2024

    work page 2024

  36. [39]

    Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark,

    Yang et al., “Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark,” in31st ACM Int. Conf. on Multimedia, pp. 4492–4501, 2023

    work page 2023

  37. [40]

    DALDA: Data Augmentation Leveraging Diffusion Model and LLM with Adaptive Guidance Scaling

    Jung et al., “Dalda: Data augmentation leveraging diffusion model and llm with adaptive guidance scaling,”arXiv:2409.16949, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  38. [41]

    Diversify your vision datasets with automatic diffusion-based augmentation,

    Dunlap et al., “Diversify your vision datasets with automatic diffusion-based augmentation,” Advances in Neural Information Pro- cessing Systems, vol. 36, 2024

    work page 2024

  39. [42]

    Effective data augmentation with diffusion models,

    Trabucco et al., “Effective data augmentation with diffusion models,”arXiv:2302.07944, 2023. Figure 10. Visual results of context experiments. Figure 11. Visual examples of dataset expansion with each dataset explored. For RAPzs [15], PETAzs [15] and PA100k [14], first row conditional image and at the second row the generated image

    Pith/arXiv arXiv 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.