Source-linked AI summary
Learning Discriminative Features with Multiple Granularities for Person Re-Identification
Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, Xi Zhou
TL;DR
Person Re-ID needs representations that remain discriminative despite difficult visual variation and the limits of global or semantically localized features. MGN learns global and uniformly partitioned local features at multiple granularities through an end-to-end multi-branch network, achieving state-of-the-art results across mainstream benchmarks, including 96.6% Rank-1 and 94.2% mAP on Market-1501 after re-ranking.
Problem
Global features can miss fine identity cues, while semantic part localization is difficult and less robust to large visual variances.
Method
MGN uses one global branch and two local branches that learn features from uniformly partitioned stripes with different numbers of parts.
Results
MGN achieves state-of-the-art performance across mainstream Re-ID datasets, including Rank-1/mAP=96.6%/94.2% on Market-1501 after re-ranking.
Takeaways & Limitations
Different branches cooperatively capture coarse global and finer local information, providing strong discriminative representations for person Re-ID.
Abstract
from arXiv · showhide
The combination of global and partial features has been an essential solution to improve discriminative performances in person re-identification (Re-ID) tasks. Previous part-based methods mainly focus on locating regions with specific pre-defined semantics to learn local representations, which increases learning difficulty but not efficient or robust to scenarios with large variances. In this paper, we propose an end-to-end feature learning strategy integrating discriminative information with various granularities. We carefully design the Multiple Granularity Network (MGN), a multi-branch deep network architecture consisting of one branch for global feature representations and two branches for local feature representations. Instead of learning on semantic regions, we uniformly partition the images into several stripes, and vary the number of parts in different local branches to obtain local feature representations with multiple granularities. Comprehensive experiments implemented on the mainstream evaluation datasets including Market-1501, DukeMTMC-reid and CUHK03 indicate that our method has robustly achieved state-of-the-art performances and outperformed any existing approaches by a large margin. For example, on Market-1501 dataset in single query mode, we achieve a state-of-the-art result of Rank-1/mAP=96.6%/94.2% after re-ranking.
1 INTRODUCTION
Person Re-ID must handle substantial variation across surveillance images, while global features can miss fine or infrequent identity cues. MGN addresses this by combining global and uniformly partitioned local features at multiple granularities, achieving strong benchmark performance.
- Person Re-ID retrieves a person from gallery images captured across different security cameras, despite pose, occlusion, clothing, clutter, and detection variations.
- Global feature learning captures salient whole-body appearance cues but can ignore non-salient details and struggle with similar inter-class properties or large intra-class differences.
- Local body-part representations complement global features by filtering outside-region distractions and concentrating learning on identity-related information.
- MGN combines global and local information across granularities, using equally split stripes rather than semantically located body regions.Increasing the number of horizontal stripes produces finer-grained discriminative responses.
- MGN achieves state-of-the-art performance on mainstream person Re-ID datasets without additional external data or re-ranking.
2 RELATED WORKS
Related Re-ID work develops deep representations, part-based feature learning, and classification or metric-loss supervision. These approaches motivate combining local stripe representations with global features, while MGN uses a multi-branch architecture for this purpose.
- Deep Re-ID research progressed from siamese and part-based networks to IDE embeddings and other methods that improve learned representations.
- Recent methods improve Re-ID through part-based alignment, vertically sliced feature maps, recurrent fusion, or direct concatenation of local features with global representations.
- MGN's architecture splits a ResNet-50 backbone into global, Part-2, and Part-3 branches after res_conv4_1.
- Part-based learning strategies use horizontal stripes, pose landmarks, region proposals, or spatial attention to obtain local representations.
- Classification losses and metric losses provide supervisory signals, with contrastive and triplet losses respectively shaping inter-class and intra-class distances.
3 MULTIPLE GRANULARITY NETWORK
MGN combines global and local representations across multiple granularities in a multi-branch, end-to-end architecture. It uses uniformly partitioned stripes, branch-specific processing, and complementary classification and metric-learning supervision.
- Network Architecture: MGN combines one global branch with Part-2 and Part-3 branches to learn pedestrian representations at different granularities.The global branch captures coarse information, while increasingly partitioned local branches target finer discriminative details.
- Network Architecture: The global branch uses down-sampling and global max-pooling, whereas local branches preserve receptive-field areas and uniformly split feature maps into horizontal stripes.Part-N Branch denotes a local branch with N partitions, including Part-2 and Part-3.
- Network Architecture: At testing, all 256-dimensional reduced features are concatenated to combine global and local information into the final representation.This concatenation is intended to improve the comprehensiveness of the learned features.
- Loss Functions: MGN trains with softmax loss for classification and batch-hard triplet loss for metric learning across its global and local embeddings.Softmax supervision is applied to specified non-reduced global and reduced local features, while triplet loss trains reduced features.
- Loss Functions: The classification-before-metric design applies softmax to coarse features and triplet loss to reduced features, supporting coarse-to-fine learning and robust convergence.The method omits triplet loss on local features because local-region contents may vary dramatically due to misalignment or related issues.
- Discussions: Separate branches outperform a single mixed branch because shared architectural stages respond to different detail levels, whereas one branch can dilute detailed information.Splitting the backbone after shallower or deeper layers also did not yield better performance in the reported experiments.
- Discussions: The global, Part-2, and Part-3 branches learn complementary coarse-to-fine preferences that cooperatively supplement discriminating information.The branch with more partitions learns finer pedestrian representations, while the global branch captures integral but coarse features.
4 EXPERIMENT
Experiments evaluate MGN across three mainstream Re-ID datasets using standard CMC and mAP protocols, showing strong performance and component effectiveness. Results include state-of-the-art comparisons, ranking examples, and ablations of architecture and losses.
- Datasets and Protocols: MGN is evaluated on Market-1501, DukeMTMC-reID, and CUHK03 using rank-1, rank-5, rank-10, and mAP metrics.Market-1501 is tested in single-query and multiple-query modes, while CUHK03 uses the protocol from [50].
- Comparison with State-of-the-Art Methods: 96.6%/94.2% Rank-1/mAP is achieved on Market-1501 in single-query mode after re-ranking, surpassing existing methods by a large margin.Without re-ranking, MGN reaches 95.7%/86.9%, exceeding PCB+RPP by 1.9% in Rank-1 and 5.3% in mAP.
- Comparison with State-of-the-Art Methods: 88.7%/78.4% Rank-1/mAP is achieved on DukeMTMC-reID, outperforming GP-reid by +3.5% in Rank-1 and +5.6% in mAP.The paper describes DukeMTMC-reID as a challenging dataset for person Re-ID.
- Comparison with State-of-the-Art Methods: 68.0%/67.4% and 66.8%/66.0% Rank-1/mAP are achieved on CUHK03 labeled and detected settings, respectively, outperforming published results by a large margin.The paper notes an obvious gap between labeled and detected conditions and links it to detection failure.
- Effectiveness of Components: 95.3%/86.2% Rank-1/mAP is obtained by MGN without triplet loss versus 87.5%/71.4% for ResNet-50, a gain of +7.8%/+14.8%.The comparison with ResNet-101 indicates that the improvement is attributed to the network architecture rather than simply additional weights.
- Effectiveness of Components: Removing Part-3 degrades Rank-1/mAP by −0.9%/0.7%, adding Part-4 gives no obvious boost, and the Part-2/Part-3 setting is reported as optimal.The Part-3/Part-4 setting performs better than Part-2/Part-4, which the authors associate with overlap between partitions.
- Effectiveness of Components: Triplet loss improves the baseline by +1.2%/3.6% and MGN by +0.4%/0.7% in Rank-1/mAP, with larger gains on mAP than Rank-1.The paper attributes the smaller MGN gain to its architecture’s enhancement of local representation.
5 CONCLUSION
MGN is a multi-branch network that learns global and local person Re-ID representations at specified body-part granularities. It directly learns local features from horizontally split stripes without part-locating operations and reports state-of-the-art results across mainstream datasets.
- MGN uses separate branches to learn global or local representations at specific body-part granularities.
- Local features are learned directly from horizontally split feature stripes in an end-to-end process.
- MGN introduces no part-locating operations such as region proposal or pose estimation.
- MGN achieves state-of-the-art results on several mainstream person Re-ID datasets.