[Paper Review] In Defense of the Classification Loss for Person Re-Identification

Yao Zhai, Xun Guo+2 more|arXiv (Cornell University)|Sep 16, 2018

Video Surveillance and Tracking Methods30 references3 citations

TL;DR

This paper proposes a channel grouping and multi-branch classification framework that significantly improves person re-identification performance using global features and standard classification loss, achieving state-of-the-art accuracy with faster inference than part-based models. By dividing global features into channel groups and training separate classification heads per group, the method enhances discriminative power while maintaining low computational cost.

ABSTRACT

The recent research for person re-identification has been focused on two trends. One is learning the part-based local features to form more informative feature descriptors. The other is designing effective metric learning loss functions such as the triplet loss family. We argue that learning global features with classification loss could achieve the same goal, even with some simple and cost-effective architecture design. In this paper, we first explain why the person re-id framework with standard classification loss usually has inferior performance compared to metric learning. Based on that, we further propose a person re-id framework featured by channel grouping and multi-branch strategy, which divides global features into multiple channel groups and learns the discriminative channel group features by multi-branch classification layers. The extensive experiments show that our framework outperforms prior state-of-the-arts in terms of both accuracy and inference speed.

Motivation & Objective

• To address the underperformance of standard classification loss in person re-identification due to a mismatch between training and testing objectives.
• To exploit the discriminative potential of global features by partitioning them into channel groups for more robust feature learning.
• To design a cost-effective architecture that matches or exceeds part-based models in accuracy while reducing inference time.
• To demonstrate that multi-branch classification on grouped global features can outperform metric learning and part-based feature learning.

Proposed method

• The method partitions the global feature map into multiple channel groups, each processed independently through a shared convolutional layer.
• Each channel group is fed into a separate fully connected classification layer, enabling multi-branch learning of discriminative features.
• During inference, features from individual channel groups are either used separately or concatenated, with no additional distance computation cost.
• The framework uses cross-entropy softmax loss for training, with no need for hard sample mining or complex metric loss functions.
• The architecture is integrated with part-based models by concatenating channel group features with part features, further boosting performance.
• The method is evaluated on Market-1501, DukeMTMC-reID, and CUHK03 benchmarks using standard training and evaluation protocols.

Experimental results

Research questions

• RQ1 Can standard classification loss achieve state-of-the-art performance in person re-identification when properly enhanced?
• RQ2 Does partitioning global features into channel groups improve discriminative feature learning compared to part-based or global feature learning?
• RQ3 Can a multi-branch classification design on channel groups match or exceed the performance of metric learning losses without increasing inference cost?
• RQ4 How does the proposed method compare to part-based models in accuracy and inference speed?

Key findings

• The proposed method achieves 93.9% rank-1 accuracy and 80.5% mAP on Market-1501, surpassing previous state-of-the-art methods.
• On DukeMTMC-reID, the method achieves 84.7% rank-1 accuracy and 69.4% mAP, setting a new SOTA.
• The framework outperforms PCB and PCB+RPP on all benchmarks while maintaining significantly faster inference speed.
• The channel grouping strategy alone improves performance over standard classification loss, and combining it with part-based features yields further gains.
• The method achieves competitive performance with minimal computational overhead, as inference cost is nearly equivalent to standard classification networks.
• Visualizations show that the learned features have larger inter-class variance and smaller intra-class variance than those from standard classification or triplet loss.