PaperlayerView exact revision on arXiv

Source-linked AI summary

DiffReID: Discriminative Diffusion Model for Object Re-Identification

Yingquan Wang, Pingping Zhang, Dong Wang, Huchuan Lu

arXiv:2609.36894v1cs.CV

TL;DR

Object ReID faces limited dataset diversity and methods that do not explicitly learn identity-aware feature distributions. DiffReID uses visual-guided diffusion to learn those distributions and mutually train guided and visual features; the authors report better results than existing works on widely used benchmarks.

  • Problem

    ReID methods face limited dataset diversity, and standard Gaussian noise lacks visual priors that could help generated features preserve identity-related information.

    Source: B. Vision-guided Noise Generator · p. 4

  • Method

    DiffReID prompt-tunes identity-aware text features, uses visual-guided noise and conditioned denoising to generate guided features, and applies mutual learning with visual features.

    Source: §I · p. 1, A. Overview of Our Framework · p. 4, D. Mutual Enhancement Constraint · p. 5

  • Results

    DiffReID performs better than existing works on several widely used object ReID benchmarks.

    Source: §V · p. 11

  • Takeaways & Limitations

    The authors conclude that mutual learning between guided and visual features improves representation generalization and discrimination.

    Source: §V · p. 11

Abstract

from arXiv · show

As a fundamental image processing task, object Re-Identification (ReID) aims to retrieve objects across non-overlapping cameras. Recently, with the development of deep learning, significant advancements have been made in object ReID. However, most existing methods suffer from generalization due to the limited size and diversity of ReID datasets. Meanwhile, current models tend to focus on extracting semantic patterns rather than learning identity-aware feature distributions. To address these issues, we propose a novel feature learning framework named \textbf{DiffReID} for object ReID. It leverages a discriminative diffusion model to gradually learn identity-aware distributions and generate identity-invariant features. More specifically, with the Contrastive Language-Image Pre-training (CLIP) model, we first obtain identity-aware text features by prompt tuning. Then, we propose a Vision-guided Noise Generator (VNG) to initialize probabilistic noises and gradually corrupt identity-aware text features. Afterwards, we take visual features as conditions and propose a Light Weight Denoiser (LWD) to denoise the corrupted text features step-by-step for identity-aware distribution learning. To obtain discriminative features, we further generate identity-invariant guided features from randomly sampling visual-guided noises. Finally, we propose a Mutual Enhancement Constraint (MEC) to facilitate mutual learning between visual features and guided features to enhance the representation robustness and discrimination. Extensive experiments on five object ReID benchmarks demonstrate that our method shows better results than most state-of-the-art methods. The source code is available at https://github.com/AWangYQ/DiffReID.

I. INTRODUCTION

Object ReID must identify objects across non-overlapping cameras, but limited dataset size and diversity constrain generalization. DiffReID applies discriminative diffusion to learn identity-aware distributions and generate identity-invariant guided features.

I. INTRODUCTION

  • Limited ReID dataset size and diversity constrain generalization, while existing methods emphasize semantic patterns over identity-aware feature distributions.ReID matches objects across non-overlapping cameras and seeks identity-invariant features despite visual variation.Source: §I · p. 1
  • Prior ReID diffusion frameworks either inject Gaussian noise across backbone layers or reuse pretrained diffusion features, with limitations in representation or inference cost.Layer-wise noise may weaken hierarchical features; complex pretrained models and large timesteps increase cost and inference time.Source: §I · p. 1
  • DiffReID uses discriminative diffusion to learn identity-aware distributions and generate identity-invariant guided features.Its VNG initializes noise, LWD generates features, and MEC refines them through mutual learning.Source: §I · pp. 1–2
  • Experiments on five object ReID benchmarks show outstanding performance for DiffReID.The introduction reports better results than most state-of-the-art methods.Source: §I · p. 2

A. Diffusion Models for Representation Learning

  • DiffReID frames diffusion as discriminative feature-distribution learning, using fine-grained conditions rather than general feature extraction or image generation.The paper states that the framework maintains efficient inference while obtaining discriminative features.Source: A. Diffusion Models for Representation… · p. 2, §I · p. 2

B. Image-based Object Re-Identification

Image-based ReID methods have expanded from CNNs and attention models to vision-language and diffusion approaches, while dataset availability and identity preservation remain concerns. DiffReID combines prompt-tuned text features, visual-guided diffusion, and mutual learning with visual features.

B. Image-based Object Re-Identification

  • Image-only ReID can overfit salient regions, and image generation may lose identity information; limited large-scale training data remains a constraint.Prior approaches include CNNs, Transformers, vision-language models, and diffusion-based data enrichment.Source: B. Image-based Object Re-Identification · p. 2

III. OUR PROPOSED METHOD

  • DiffReID uses CLIP prompt tuning to obtain identity-aware text features, then VNG corrupts them and LWD denoises them for distribution learning.MEC facilitates mutual learning between visual and guided features to improve robustness and discrimination.Source: §III · pp. 2–3, Figure 2 · p. 3

A. Overview of Our Framework

  • Training first freezes CLIP's text and image encoders while optimizing prompt tokens with bidirectional contrastive losses.The prompt template is “A photo of a [x] [x] [x] [x] person/vehicle”; P(y) denotes images of identity y, and B is batch size.Source: A. Overview of Our Framework · p. 3
  • In stage two, visual features guide noise initialization and condition reverse denoising, which progressively converts corrupted text features into guided features.The process operates in latent feature space on identity-aware text features, avoiding pixel-level reconstruction while using visual features as conditions.Source: A. Overview of Our Framework · pp. 3–4

B. Vision-guided Noise Generator

The Vision-guided Noise Generator uses visual features to initialize diffusion noise, so the starting state can carry image-specific priors rather than relying on sample-agnostic Gaussian noise.

B. Vision-guided Noise Generator

  • Standard Gaussian initialization is sample-agnostic and lacks visual priors, so its perturbations may conflict with object identity and reduce generated-feature discrimination.The noise initialization determines the denoising process's starting state and influences identity preservation.Source: B. Vision-guided Noise Generator · p. 4
  • VNG estimates a noise distribution's mean and variance from the visual feature, then samples initial visual-guided noise from that learned distribution.Two fully connected layers transform the visual feature into the mean and variance; sampling uses the reparameterization trick.Source: B. Vision-guided Noise Generator · p. 4
  • KL-divergence regularization constrains the learned visual-conditioned noise distribution toward a standard Gaussian.The visual features serve as priors while identity-aware text features are gradually corrupted.Source: B. Vision-guided Noise Generator · p. 4

C. Light Weight Denoiser

The Light Weight Denoiser (LWD) denoises identity-aware text features in a compact latent space using visual conditions. Timestep-specific conditions adapt the guidance as features move from noisy structure recovery to finer identity discrimination.

C. Light Weight Denoiser

  • LWD uses MLPs in a compact latent feature space to denoise semantic representations with computational efficiency.The framework stacks M denoisers; the paper states that an MLP-based denoiser with a small number of timesteps is sufficient.Source: C. Light Weight Denoiser · p. 4
  • Intermediate image-encoder features are aggregated and linearly projected into the visual condition representation used by LWD.The noised feature, time embedding, and timestep-specific condition are jointly processed to denoise corrupted text features.Source: C. Light Weight Denoiser · pp. 4–5
  • Temporal Condition Queries provide timestep-specific visual guidance as denoising shifts from coarse identity structure to fine-grained discriminative cues.The paper motivates distinct conditions because early features contain stronger noise, while later features are progressively refined.Source: C. Light Weight Denoiser · p. 4
  • At each timestep, learnable queries incorporate the current noisy feature as a prior and interact with the visual condition representation through cross-attention.The queries are initialized per timestep; the noise prior is described as stabilizing training and accelerating convergence.Source: C. Light Weight Denoiser · p. 5

D. Mutual Enhancement Constraint

The Mutual Enhancement Constraint (MEC) combines visual and generated guided features in a joint set for cross-feature triplet learning. This lets the two feature types refine one another while addressing their complementary properties.

D. Mutual Enhancement Constraint

  • Generated features provide identity-aware semantic cues, while visual features constrain unreliable semantic deviations with image-specific evidence.The paper contrasts visual features, which capture discriminative information but tend to overfit salient appearance patterns, with generated features encouraged to follow the identity-aware distribution.Source: D. Mutual Enhancement Constraint · p. 5
  • MEC combines visual and guided features in a joint set and uses cross-feature triplets to pull same-identity pairs together and push different-identity features apart.The visual feature acts as a stable identity reference for guided features, while guided-feature variation encourages identity-consistent cues in visual features.Source: D. Mutual Enhancement Constraint · p. 5

E. Training and Inference

Training uses prompt tuning followed by a second stage that supervises visual, diffusion, and guided features with multiple losses. At inference, multiple guided features are averaged into the final feature.

E. Training and Inference

  • The first training stage prompt-tunes using bidirectional contrastive losses, and the second stage optimizes the visual feature with identity and triplet losses.The second-stage visual loss combines the identity loss and triplet loss with hyperparameters balancing the terms.Source: E. Training and Inference · p. 5
  • The second stage also supervises diffusion noise prediction and guided-feature identity, and includes a reformulated KL divergence term in its overall loss.The stated overall loss comprises visual, diffusion, MEC, discrimination, and KL terms.Source: E. Training and Inference · p. 5
  • Table I compares state-of-the-art methods on three person ReID datasets and one vehicle ReID dataset.The caption says best and second-best results are marked in bold and underlined, respectively.Source: Table I · p. 6
  • At inference, the final feature is the average of guided features generated from visual-guided noises.The paper averages multiple samples because a single guided feature may reflect only one possible representation of the learned identity-aware distribution.Source: E. Training and Inference · p. 6

IV. EXPERIMENTS

The experiments evaluate object ReID on five datasets using two domain-generalization settings. Performance is assessed with mAP and Rank-1 CMC.

IV. EXPERIMENTS

B. Implementation Details

DiffReID outperforms compared methods in single-domain and domain-generalization evaluations, while ablations examine its components, diffusion design, and training robustness. Implementation uses a two-stage training setup with a pretrained CLIP backbone.

B. Implementation Details

  • Implementation uses pretrained CLIP for image and text feature extraction and two training stages, with batch size 64 and 60 epochs in stage two.The image and text inputs use different resize dimensions; stage one optimizes learnable prompt tokens without augmentation.Source: B. Implementation Details · p. 6

C. Comparison with State-of-the-Arts Methods

D. Ablation Studies

  • With all components, DiffReID reaches 79.0% mAP on MSMT17 and 85.6% mAP on DukeMTMC; VNG and MEC add further gains.VNG adds 1.1% and 0.6% mAP on MSMT17 and DukeMTMC, while MEC adds 0.6% and 0.7%, respectively.Source: Table III · p. 7, D. Ablation Studies · p. 7
  • With T = 10 diffusion steps, the model yields its best performance; even T = 2 reaches 77.0% mAP on MSMT17 and 84.4% mAP on DukeMTMC.Performance remains competitive from T = 4 to T = 8.Source: Table IV · p. 7, D. Ablation Studies · p. 7
  • MEC improves performance for both visual and guided features, while increasing intra-class similarity and generation consistency and decreasing inter-class similarity.The authors interpret these findings as evidence that MEC produces more discriminative and stable guided features.Source: D. Ablation Studies · pp. 7–8, Table V · p. 8
  • Three guided features balance performance and efficiency; 16 temporal queries perform best; intermediate features help, but too many hurt, and results remain stable across random seeds.Across four seeds, mAP ranges from 78.4% to 79.0% and Rank-1 from 90.7% to 91.4%.Source: D. Ablation Studies · p. 8
  • On MSMT17, visual-feature diffusion raises mAP from 75.4% to 77.8% and Rank-1 from 89.3% to 90.5%; identity-aware text diffusion reaches 79.0% mAP and 91.2% Rank-1.With KL regularization, performance is 79.0% mAP and 91.2% Rank-1; without it, results are 77.6% mAP and 90.8% Rank-1.Source: D. Ablation Studies · pp. 8–9, Table X · p. 9, Table XI · p. 9

E. Computational Cost Analysis

DiffReID uses an MLP-based Light Weight Denoiser instead of a complex U-Net. Its computational overhead is comparable to other methods while its ReID performance is reported as superior.

E. Computational Cost Analysis

  • The MLP-based LWD is more efficient than previous diffusion-based methods such as SD-ReID, with comparable computational overhead and superior ReID performance overall.The authors describe this as a trade-off between computational cost and performance.Source: Table XII · p. 9, E. Computational Cost Analysis · p. 9

F. Qualitative Analysis

Qualitative analyses show DiffReID retrieving correct matches despite distracting attributes and distortions, while its features become more identity-separated and discriminative through denoising and visual-guided noise.

Top-5 Retrieval Results

  • DiffReID retrieves correct matches despite prominent visual distractions and query distortions that mislead CLIP-ReID.The examples include similar clothing, a white backpack, and a distorted query.Source: F. Qualitative Analysis · p. 9

Visualization of Feature Distribution

  • On MSMT17, DiffReID reduces intra-class distances and increases inter-class distances compared with the baseline.The t-SNE visualization samples twenty identities from the test set.Source: F. Qualitative Analysis · p. 9, Figure 5 · p. 10
  • The cosine-similarity distribution gap increases from 0.5 to 0.598 with visual-guided noise, while negative pairs shift lower and positive pairs remain highly similar.This indicates greater inter-identity separation while preserving intra-identity consistency.Source: F. Qualitative Analysis · p. 9, Figure 6 · p. 10

Visualization of Guided Features Across the Diffusion Process

  • Across denoising timesteps, guided features evolve from overlapping, high-variance representations into compact, well-separated identity clusters.The feature shift between h = 6 and h = 7 reflects recovery from low-frequency global structure to high-frequency identity cues, not instability.Source: F. Qualitative Analysis · p. 10, Figure 4 · p. 10

Visual Effect of TCQ

  • Attention shifts from global person structure toward discriminative regions over denoising steps and suppresses irrelevant content, including an unrelated person.Darker red in the attention maps indicates higher attention weights.Source: F. Qualitative Analysis · pp. 10–11, Figure 7 · p. 10

V. CONCLUSION

DiffReID learns identity-aware distributions and generates guided features to complement visual features, with VNG, LWD, and MEC supporting representation learning. The authors report better performance than existing works on several widely used object ReID benchmarks.

V. CONCLUSION

  • DiffReID combines VNG, LWD, and MEC to learn identity-aware distributions and generate diverse guided features that complement visual features for robust representation.The authors report that extensive experiments show better performance than existing works on several widely used object ReID benchmarks.Source: §V · p. 11
Loading 2609.36894v1…