[Paper Review] Auto-ReID: Searching for a Part-aware ConvNet for Person Re-Identification

Ruijie Quan, Xuanyi Dong+3 more|arXiv (Cornell University)|Mar 23, 2019

Video Surveillance and Tracking Methods45 references42 citations

TL;DR

Auto-ReID automatically searches for a part-aware CNN architecture tailored for person re-ID, introducing a reID-specific search space and retrieval-based NAS to achieve state-of-the-art results with fewer parameters.

ABSTRACT

Prevailing deep convolutional neural networks (CNNs) for person re-IDentification (reID) are usually built upon ResNet or VGG backbones, which were originally designed for classification. Because reID is different from classification, the architecture should be modified accordingly. We propose to automatically search for a CNN architecture that is specifically suitable for the reID task. There are three aspects to be tackled. First, body structural information plays an important role in reID but it is not encoded in backbones. Second, Neural Architecture Search (NAS) automates the process of architecture design without human effort, but no existing NAS methods incorporate the structure information of input images. Third, reID is essentially a retrieval task but current NAS algorithms are merely designed for classification. To solve these problems, we propose a retrieval-based search algorithm over a specifically designed reID search space, named Auto-ReID. Our Auto-ReID enables the automated approach to find an efficient and effective CNN architecture for reID. Extensive experiments demonstrate that the searched architecture achieves state-of-the-art performance while reducing 50% parameters and 53% FLOPs compared to others.

Motivation & Objective

• Motivate the need for reID-specific architectural designs beyond classification backbones.
• Propose a novel reID-oriented search space that encodes body-part structure as a trainable module.
• Develop a retrieval-based NAS algorithm that optimizes for reID objectives rather than classification only.
• Demonstrate that the searched architecture achieves superior accuracy with reduced model size and computation.

Proposed method

• Design a reID search space by integrating a novel part-aware module that captures body structure information.
• Adopt a differentiable NAS framework to optimize architecture parameters (alpha) and operation parameters (omega) under a retrieval loss.
• Use a macro backbone based on a ResNet-like structure with neural cells whose topology is searched.
• Introduce a mixture retrieval loss L_ret = lambda L_s + (1 - lambda) L_t to align NAS with reID objectives.
• Train and evaluate the final searched CNN on standard reID benchmarks with and without ImageNet pre-training.

Experimental results

Research questions

• RQ1 Can neural architecture search yield reID-specific architectures that better capture body structure than classification-backed backbones?
• RQ2 Does incorporating a trainable part-aware module within the NAS search space improve reID performance and efficiency?
• RQ3 Can a retrieval-oriented NAS objective outperform traditional classification-based NAS objectives for reID tasks?
• RQ4 What are the performance and efficiency gains (parameters, FLOPs) of the Auto-ReID architecture on standard reID benchmarks?

Key findings

• The Auto-ReID search space with a part-aware module yields architectures with competitive or superior reID performance while having significantly fewer parameters and FLOPs.
• Compared to baselines, the Auto-ReID searched model achieves higher rank-1 and mAP on Market-1501 and shows strong transfer to CUHK03 and MSMT17.
• With ImageNet pre-training, Auto-ReID attains state-of-the-art performance on Market-1501 and surpasses several hand-crafted backbones while reducing parameters by more than 45%.
• Using the retrieval loss in NAS improves CNN quality for reID in most configurations tested.
• In ablations, the reID-specific search space and retrieval-based search outperform Baseline, Baseline+ReID Search Space, and DARTS/GDAS baselines.