PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Group-shuffling Random Walk for Person Re-identification

Yantao Shen, Hongsheng Li, Tong Xiao, Shuai Yi, Dapeng Chen, Xiaogang Wang

arXiv:1807.11178v1cs.CV

TL;DR

Person re-identification ranks gallery images by their affinities to a probe, but those direct affinities can be insufficient when views differ. The paper introduces an end-to-end group-shuffling random walk that uses gallery-to-gallery affinities, and reports that it outperforms baseline and state-of-the-art approaches by large margins.

  • Problem

    Relying only on probe-to-gallery affinities may be insufficient for ranking images of the same person, while gallery-to-gallery affinities can help determine the final rankings.

    Source: §1 · p. 1

  • Method

    A group-shuffling random-walk layer uses initial P2G and G2G affinities in end-to-end training and testing, with grouped features and shuffled affinity pairs providing additional supervision.

    Source: §1 · p. 2

  • Results

    The proposed approach outperforms baseline methods and state-of-the-art approaches by large margins.

    Source: §5 · p. 8

  • Takeaways & Limitations

    Grouping and shuffling provide rich supervision on P2G and G2G affinities and better regularize feature learning.

    Source: §1 · p. 2

Abstract

from arXiv · show

Person re-identification aims at finding a person of interest in an image gallery by comparing the probe image of this person with all the gallery images. It is generally treated as a retrieval problem, where the affinities between the probe image and gallery images (P2G affinities) are used to rank the retrieved gallery images. However, most existing methods only consider P2G affinities but ignore the affinities between all the gallery images (G2G affinity). Some frameworks incorporated G2G affinities into the testing process, which is not end-to-end trainable for deep neural networks. In this paper, we propose a novel group-shuffling random walk network for fully utilizing the affinity information between gallery images in both the training and testing processes. The proposed approach aims at end-to-end refining the P2G affinities based on G2G affinity information with a simple yet effective matrix operation, which can be integrated into deep neural networks. Feature grouping and group shuffle are also proposed to apply rich supervisions for learning better person features. The proposed approach outperforms state-of-the-art methods on the Market-1501, CUHK03, and DukeMTMC datasets by large margins, which demonstrate the effectiveness of our approach.

1. Introduction

Person re-identification ranks gallery images by their probe-to-gallery affinities, but gallery-to-gallery relationships can help when direct probe-gallery matching is weak. The paper proposes an end-to-end group-shuffling random walk that uses those relationships while learning person features.

1. Introduction

  • Person re-identification is framed as retrieving gallery images ranked by their affinities to a probe image.Source: §1 · p. 1
  • G2G affinities can help match a person’s frontal and back views when a side-view image has high affinity with both.Source: §1 · p. 1
  • Earlier re-ranking methods used gallery neighbors or manifold ranking, but operated as separate post-processing rather than improving feature learning during training.Source: §1 · pp. 1–2
  • The group-shuffling random walk integrates P2G and G2G affinities end-to-end, refining rankings while learning more discriminative person features.Unlike separate re-ranking, it can be trained within a deep network.Source: Figure 1 · p. 1, §1 · p. 2
  • Feature grouping applies supervision separately to feature dimensions, while group shuffling combines pairs of grouped P2G and G2G affinities to enrich supervision and regularize feature learning.Source: §1 · p. 2

2. Related Work

Related work spans deep person-feature learning, gallery-based re-ranking, and random-walk methods. The paper’s approach uses inter-image affinities to refine probe-gallery rankings within a deep network.

2. Related Work

  • Deep person re-identification methods use CNNs to learn features, including approaches for misalignment, pairwise relationships, metric learning, domain gaps, and pose variation.Metric-learning approaches also learn distance measures for person-image similarity.Source: §2 · p. 2
  • Prior re-ranking methods incorporated gallery affinities through human interaction, local and global features, or k-reciprocal neighbors.Source: §2 · p. 2
  • Earlier random-walk methods refined person re-identification rankings as testing-time post-processing, rather than end-to-end training that can improve feature learning.Source: §2 · p. 2
  • The proposed pipeline estimates initial P2G and G2G affinities with a pairwise CNN, then applies feature grouping and group-shuffling random walk to refine P2G affinities.Source: Figure 2 · p. 3
  • Random walks were also used for image segmentation through inter-pixel relations, whereas this paper focuses on inter-image relations for affinity ranking.Source: §2 · pp. 2–3

3. Approach

The approach embeds random-walk refinement in an end-to-end network, using gallery-to-gallery affinities to refine probe-to-gallery scores. Feature grouping and shuffling provide additional supervision for learning person features.

3.1. Random walk algorithm

  • Gallery images form a graph whose normalized pairwise affinities populate W, and repeated multiplication y(t+1)=Wy(t) diffuses G2G information into P2G scores.Source: §3.1 · p. 3

3.2. Random walk in deep neural networks

  • A Siamese CNN estimates initial P2G and G2G affinities; the layer propagates scores, combines updates with initial estimates, and produces end-to-end trainable refined affinities.The G2G affinity matrix is row-normalized, with diagonal entries set to zero to prevent self-reinforcement.Source: §3.2 · pp. 3–4

3.3. Rich supervisions from random walk

  • Unlike conventional Siamese CNNs, whose single-pair loss updates only that pair’s P2G affinity, random walk backpropagates the loss through all P2G and G2G affinities and image features.Source: §3.3 · p. 4

3.4. Group-shuffling random walk

  • Features are split into K sub-vectors and their groupwise P2G and G2G affinities are pairwise recombined for random walks, giving each feature group separate supervision.The grouping addresses the risk that some feature dimensions remain inactive when other dimensions dominate training.Source: §3.4 · p. 5

3.5. Overall network structure

  • The overall network combines a ResNet-50 Siamese affinity CNN with a group-shuffling random-walk layer, producing K affinity scores per image pair and K^2 refined P2G scores.Cross-entropy supervision is applied to the refined affinities.Source: §3.5 · pp. 5–6

4. Experiments

Experiments and ablations evaluate the method on Market-1501, CUHK03, and DukeMTMC, examining its components and comparisons with prior methods. The reported final results are strongest overall, with a CUHK03 top-10 exception.

4.1. Datasets and metric

  • The experiments use Market-1501, CUHK03, and DukeMTMC, reporting mAP and CMC top-1, top-5, and top-10 accuracy under each dataset’s original computation setup.The datasets contain person images cropped from market or campus scenes; Market-1501 uses detector crops, while CUHK03 and DukeMTMC use manual crops.Source: §4.1 · p. 6

4.2. Implementation details

  • Training uses 256-image batches, with one probe and three gallery images per identity; testing reranks the top 75 gallery images using averaged affinities from four groups.The final model uses K = 4 and λ = 0.95; the network generally converges after 100 epochs.Source: §4.2 · p. 6

4.3. Ablation study

  • With the same ResNet-50 structure, the verification-loss baseline outperforms improved triplet loss by 13.7% mAP.The baseline uses K = 1 and supervises all 64 × 192 P2G pairs and 192² G2G pairs in each mini-batch.Source: §4.3 · p. 6
  • Feature grouping outperforms dropout on Market-1501 and CUHK03, and combining grouping with dropout improves further; the proposed operations also contribute to better visual features.Dropout ratios were 0.5 for Market-1501 and 0.3 for CUHK03; the feature analysis uses trained ResNet-50 distances on Market-1501 and DukeMTMC.Source: §4.3 · pp. 6–7
  • End-to-end random walk outperforms both k-reciprocal re-ranking and a separate post-processing random walk, supporting training the network with the walk operation.The post-processing random walk itself outperforms k-reciprocal re-ranking, but remains below the end-to-end approach.Source: §4.3 · pp. 6–7
  • End-to-end random walk raises mAP by 2.9%, 1.3%, and 2.7% on Market-1501, CUHK03, and DukeMTMC; four-way grouping adds 1.4% on CUHK03 but only marginally elsewhere.Four feature groups generally outperform no grouping by approximately 1% mAP.Source: §4.3 · p. 7
  • With K = 4, group-shuffling generally adds approximately 1% mAP across all three datasets.The comparison contrasts rows using K = 4 group-shuffling with the corresponding non-shuffled variants.Source: §4.3 · p. 7

4.4. Comparison with state-of-the-art methods

  • Across Market-1501 and DukeMTMC, the method beats all compared methods; on CUHK03 it reports 94.0% mAP and 94.9% top-1, while PDC leads top-10.The reported DukeMTMC gains over SVDNet are 9.6% mAP and 14.0% top-1.Source: §4.4 · pp. 7–8, Table 5 · p. 8, Table 6 · p. 8

5. Conclusion

The paper uses gallery-to-gallery affinities to refine probe-to-gallery affinities through group-shuffling random walk, and combines this with feature grouping and shuffling for richer supervision. Its overall performance surpasses baseline and state-of-the-art approaches by large margins.

  • The overall approach outperforms baseline and state-of-the-art methods by large margins.Source: §5 · p. 8
  • Group-shuffling random walk uses gallery-to-gallery affinities to refine probe-to-gallery affinities.Source: §5 · p. 8
  • Grouping and shuffling features provides richer supervision for learning discriminative person features.Source: §5 · p. 8
Loading 1807.11178v1…