[Paper Review] AlignedReID: Surpassing Human-Level Performance in Person Re-Identification

Xuan Zhang, Hao Luo+7 more|arXiv (Cornell University)|Nov 22, 2017

Video Surveillance and Tracking Methods39 references437 citations

TL;DR

The paper introduces AlignedReID, a joint global-local feature framework with automatic local part alignment via a shortest-path, plus mutual learning, achieving state-of-the-art results and surpassing human performance on Market1501 and CUHK03.

ABSTRACT

In this paper, we propose a novel method called AlignedReID that extracts a global feature which is jointly learned with local features. Global feature learning benefits greatly from local feature learning, which performs an alignment/matching by calculating the shortest path between two sets of local features, without requiring extra supervision. After the joint learning, we only keep the global feature to compute the similarities between images. Our method achieves rank-1 accuracy of 94.4% on Market1501 and 97.8% on CUHK03, outperforming state-of-the-art methods by a large margin. We also evaluate human-level performance and demonstrate that our method is the first to surpass human-level performance on Market1501 and CUHK03, two widely used Person ReID datasets.

Motivation & Objective

• Motivate robust person re-identification under pose, occlusion, and misalignment by leveraging both global and local cues.
• Develop an end-to-end learning framework that aligns local parts without extra supervision or pose estimation.
• Improve metric learning through mutual learning between models to boost representation quality.
• Evaluate on standard ReID datasets and compare with human performance to establish benchmarks.

Proposed method

• Use a CNN to extract a feature map.
• Compute a global feature via global pooling.
• Compute H local features via horizontal pooling plus a 1x1 conv to reduce channels.
• Define local distance as the shortest path distance on a distance matrix between local features (dynamic programming).
• Combine global and local distances for a joint metric learning loss (TriHard).
• Apply mutual learning by training two models jointly with metric and classification mutual losses; use zero-gradient sharing to stabilize training.

Experimental results

Research questions

• RQ1 Can implicit local feature alignment improve global feature learning for ReID without explicit pose supervision?
• RQ2 Does mutual learning further enhance representation quality and ranking performance beyond a single AlignedReID model?
• RQ3 To what extent can the global feature alone, aided by learned local alignment, compete with joint global+local representations at inference?
• RQ4 How does AlignedReID perform relative to human performance on standard ReID benchmarks?
• RQ5 What is the impact of re-ranking on final performance when using AlignedReID features?

Key findings

• AlignedReID substantially improves over baselines, with notable gains in rank-1 and mAP across Market1501, CUHK-SYSU, and CUHK03.
• The global feature alone, learned with local feature alignment, nearly matches or exceeds the performance of using both global and local features at inference.
• Mutual learning further boosts performance, especially when using an architecture with distinct backbones (e.g., ResNet50 vs. ResNet50-Xception).
• With re-ranking, AlignedReID achieves 94.4% rank-1 and 90.7% mAP on Market1501, and 97.8% rank-1 on CUHK03, surpassing prior methods.
• Human evaluation shows the best annotators reach 93.5% (Market1501) and 95.7% (CUHK03) rank-1 accuracy, while AlignedReID with re-ranking achieves higher on Market1501 (94.4%) and CUHK03 (97.8%).
• The method demonstrates strong generalization across datasets and reports qualitative alignment examples illustrating robustness to misalignment and occlusion.