Learning Discriminative Features with Multiple Granularities for Person Re-Identification (1804.01438v3)

Abstract: 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.

• The paper introduces a multi-granularity network (MGN) that integrates global and local features to enhance person re-identification accuracy.
• The approach utilizes three ResNet-50 based branches with uniform horizontal partitions, optimizing softmax and triplet loss for robust metric learning.
• The method achieves state-of-the-art Rank-1 and mAP scores on Market-1501, DukeMTMC-reID, and CUHK03, demonstrating significant improvements over previous methods.

Learning Discriminative Features with Multiple Granularities for Person Re-Identification

The paper "Learning Discriminative Features with Multiple Granularities for Person Re-Identification" by Wang et al., presents a novel approach leveraging multi-granularity feature learning to enhance the performances of person re-identification (Re-ID) models. The approach introduces Multiple Granularity Network (MGN), a deep multi-branch network designed to capture both global and local features at varying granularities, thus addressing challenges presented by complex image variations in surveillance tasks.

Overview

Person re-identification aims to recognize individuals across different camera views in a surveillance system. Traditional methods struggle with high variations due to pose, occlusion, and background clutter. Previous part-based approaches often relied on pre-defined body regions to extract features, which could be brittle under challenging conditions. Instead, this paper proposes MGN, which eschews semantic-based regions in favor of uniform horizontal partitions at multiple scales, thus simplifying the learning process and enhancing robustness.

Methodology

MGN consists of three branches derived from a shared ResNet-50 backbone:

1. Global Branch captures overall feature representations by down-sampling and global pooling.
2. Part-2 Branch uniformly partitions feature maps into two horizontal stripes.
3. Part-3 Branch uniformly partitions feature maps into three horizontal stripes.

Each branch has its specific pooling and reduction layers, and outputs 256-dimensional features. These features are concatenated to form a comprehensive representation capturing diverse granularity levels.

To train MGN, the authors employ a combination of softmax loss for classification and triplet loss for metric learning, enhancing both the discriminative power and distance metrics between features. A novel classfication-before-metric architecture is utilized where softmax losses are applied before the dimensionality reduction, and triplet losses are applied after.

Results

The efficacy of MGN was validated on three prominent Re-ID datasets: Market-1501, DukeMTMC-reID, and CUHK03. The observed results include:

• Market-1501: Achieved Rank-1/mAP scores of 95.7%/86.9% (single-query mode), and 96.9%/90.7% (multiple-query mode). Post re-ranking, these scores improved to 96.6%/94.2% and 97.1%/95.9% respectively.
• DukeMTMC-reID: Achieved Rank-1/mAP scores of 88.7%/78.4%.
• CUHK03: Under the new protocol, achieved Rank-1/mAP scores of 68.0%/67.4% (labeled) and 66.8%/66.0% (detected).

The MGN approach consistently outperformed existing state-of-the-art methods, notably surpassing the results of PCB+RPP $Sun et al., 2017$ and others by a considerable margin.

Implications

The architecture's simplicity and effectiveness make MGN highly applicable in practical surveillance scenarios where varied conditions and high complexity are common. The uniform splitting strategy avoids the pitfalls of pre-defined semantic regions, allowing the network to adapt to finer discriminative cues inherently present in the data.

Future Directions

Future research could investigate the effects of using different base networks or extending the multi-branch approach to consider vertical or dynamic partitions based on content. Moreover, exploring the integration of additional context cues, such as temporal sequences, may further enhance the model's robustness and adaptability to real-world scenarios.

In summary, the paper by Wang et al. presents a significant contribution to person re-identification by introducing a robust, multi-granularity network that advances the state-of-the-art through a comprehensive and straightforward multi-branch architecture.