[论文解读] Deeply-Learned Part-Aligned Representations for Person Re-Identification

Li-Ming Zhao, Xi Li+2 more|arXiv (Cornell University)|Jul 23, 2017

Video Surveillance and Tracking Methods参考文献 32被引用 166

一句话总结

论文引入一种部分对齐表示，用端到端学习人体部位特定特征，无需部位标签，提升对姿态和空间错位的鲁棒性，并在标准数据集上达到最先进的结果。

ABSTRACT

In this paper, we address the problem of person re-identification, which refers to associating the persons captured from different cameras. We propose a simple yet effective human part-aligned representation for handling the body part misalignment problem. Our approach decomposes the human body into regions (parts) which are discriminative for person matching, accordingly computes the representations over the regions, and aggregates the similarities computed between the corresponding regions of a pair of probe and gallery images as the overall matching score. Our formulation, inspired by attention models, is a deep neural network modeling the three steps together, which is learnt through minimizing the triplet loss function without requiring body part labeling information. Unlike most existing deep learning algorithms that learn a global or spatial partition-based local representation, our approach performs human body partition, and thus is more robust to pose changes and various human spatial distributions in the person bounding box. Our approach shows state-of-the-art results over standard datasets, Market-$1501$, CUHK$03$, CUHK$01$ and VIPeR.

研究动机与目标

• Address body-part misalignment in person re-identification by learning discriminative body parts directly from data.
• Propose a part-aligned representation that partitions the body into regions without labeled part data.
• Integrate part detection and feature extraction in a single deep model trained with triplet loss.
• Demonstrate robustness to pose variation and bounding-box misalignment across standard benchmarks.

提出的方法

• Use a fully convolutional network to extract image feature maps from person bounding boxes.
• Introduce a part net with multiple branches to detect discriminative body-region maps from the feature maps.
• Compute part-specific features by weighting and pooling over each detected region and reduce to fixed-length vectors.
• Concatenate and L2-normalize part features to form a global part-aligned representation.
• Train the network end-to-end with triplet loss to enforce relative distances between positives and negatives.
• Apply a simple Euclidean distance for matching between probe and gallery images, enabling efficient retrieval.

实验结果

研究问题

• RQ1 Can a part-aligned representation learned without explicit body-part labels outperform traditional spatial partitions in person re-identification?
• RQ2 Do learned body parts align well across images of the same person under pose and view changes?
• RQ3 How does the proposed part-aligned approach compare to headway methods using fixed spatial partitions or external part segmentation?
• RQ4 What is the impact of the number of parts on re-identification performance?
• RQ5 Is the method effective across multiple standard datasets and feature backbones?

主要发现

• The part-aligned representation achieves competitive to state-of-the-art performance on Market-1501, CUHK03, CUHK01, and VIPeR."
• Increasing the number of parts up to eight generally improves rank-1 accuracy, with gains stabilizing beyond that point.
• Using learned body parts outperforms spatial stripe and grid partitions in key benchmarks.
• Replacing the part net with fixed FC or pooling baselines degrades performance, showing the benefit of adaptive part detection.
• The method provides gains across different backbone networks (AlexNet, VGGNet, GoogLeNet).
• Compared with state-of-the-art methods, the proposed approach achieves the best reported results on several settings, including Market-1501 and CUHK03 (both labeled and detected boxes).