Source-linked AI summary
Auto-ReID: Searching for a Part-aware ConvNet for Person Re-Identification
Ruijie Quan, Xuanyi Dong, Yu Wu, Linchao Zhu, Yi Yang
TL;DR
Person re-identification models often reuse classification backbones despite task-specific differences and missing body-structure modeling. Auto-ReID searches a reID-specific space with part-aware components and retrieval loss, achieving state-of-the-art performance on three benchmarks with half the parameters.
Problem
Classification-oriented CNN backbones differ from reID needs, while existing NAS methods do not incorporate body structure or retrieval objectives.
Method
Auto-ReID combines a part-aware reID search space with a retrieval-based loss to automatically search CNN architectures suited to reID.
Results
Auto-ReID significantly outperforms state-of-the-art reID models on three benchmarks, while the searched CNN uses half the parameters.
Takeaways & Limitations
Auto-ReID provides an automated alternative to manual reID architecture design while explicitly incorporating pedestrian body-part information.
Takeaways & Limitations
The search space considers only one possible reID-specific module, and the search algorithm is a simple extension of existing NAS.
Abstract
from arXiv · showhide
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.
1. Introduction
Auto-ReID addresses the mismatch between classification-oriented CNNs and person re-identification by automatically searching a reID-specific architecture. It combines body-structure-aware components with retrieval-based NAS and achieves competitive or state-of-the-art accuracy with substantially fewer parameters.
- Motivation: ReID differs from classification because all inputs are person images whose attributes vary, so classification backbones should be modified for retrieval.Examples of varying attributes include apparel and hair styles.
- Motivation: Manual design of reID architectures is inefficient and labor intensive, while existing NAS methods do not preserve body structural information or directly target retrieval.The paper identifies backbone dependence and retrieval-specific search as additional challenges.
- Approach: Auto-ReID introduces a reID search space combining typical NAS operations with a trainable part-aware module that captures pedestrian body-part information.The module handles various input feature shapes and serves as a basic operation for candidate architectures.
- Approach: Auto-ReID integrates a retrieval loss into differentiable NAS and modifies searching and batch sampling to fit the reID objective.The retrieval objective supervises architecture search on a specific reID dataset.
- Results: State-of-the-art performance on three reID benchmarks is achieved with only half the number of parameters, while accuracy remains competitive against reID baselines.The searched CNN has less than 40% of baseline parameters before ImageNet pre-training.
2. Related Work
Prior reID methods improve deep CNN representations using losses, pose or joint information, and part-based features, but generally rely on classification backbones. NAS research largely targets classification, often uses proxy tasks, and does not systematically encode reID-specific information.
- Person reID: Deep reID methods enhance representations through multi-dataset features, quadruplet loss, body joint maps, or part-based pooling.These approaches target interclass and intra-class variation, pose sensitivity, or discriminative part-informed features.
- Person reID: Existing deep reID algorithms commonly rely on VGG, Inception, or ResNet backbones designed and evaluated for image classification.The paper notes that these backbones may not align with reID.
- Neural Architecture Search: Most NAS methods search on small proxy tasks and transfer architectures, while reinforcement-learning approaches can require more than hundreds of GPU days.The paper instead focuses on searching for a high-performance reID model.
- Neural Architecture Search: Generic NAS approaches target classification and do not guarantee suitability for reID without considering semantics, occlusion, pose, or body-part information.Auto-ReID discards the proxy paradigm and directly searches on the target reID dataset.
3. Methodology
Auto-ReID adapts differentiable NAS to person re-identification by searching neural-cell architectures with a retrieval objective and a part-aware search space. The method combines body-structure modeling with standard NAS operations and trains the resulting architecture through a reID pipeline.
- NAS Preliminaries: Each neural cell is a directed acyclic graph whose blocks select two input tensors, apply two candidate operations, and sum the results.Candidate inputs come from the previous two cells and the preceding block in the current cell; operation choices are relaxed with a softmax during search.
- ReID Search Algorithm: The search objective replaces classification-only optimization with a retrieval loss that mixes cross-entropy and triplet losses.Triplet-loss optimization uses a class-balance sampler because the loss depends on the identities and images included in each batch.
- Network Structure: The backbone follows ResNet’s macro structure, replacing residual layers with searched neural cells and mapping backbone features through embedding and classification layers.Two dropout layers are inserted between the backbone feature, embedding, and logit transformations.
- ReID Search Space with Part-Aware Module: The part-aware module vertically splits an input feature tensor into M = 4 parts, transforms pooled part vectors, exchanges information through self-attention, and fuses the enhanced tensor with the original.Repeated and concatenated body vectors recover the input spatial shape before one-by-one convolutional fusion.
- ReID Search Algorithm: Auto-ReID searches neural-cell topologies for reID using a bi-level differentiable NAS procedure that alternates updates to architecture and operation parameters.The final CNN is derived from the learned architecture parameters, then trained and evaluated using standard reID procedures.
- ReID Search Space with Part-Aware Module: The reID search space combines a part-aware module with pooling, depth-wise separable convolution, dilated convolution, zero, and identity operations.This extends standard NAS operations with a trainable component designed to capture pedestrian body-part information.
4. Experiments
The experiments evaluate Auto-ReID through implementation settings, ablations, and comparisons across three person re-identification benchmarks. Results show that its searched architecture combines strong accuracy with lower computational cost and transfers beyond the search dataset.
- Experimental setup: The experiments cover dataset protocols, implementation settings, ablations, state-of-the-art comparisons, and qualitative analysis.Evaluation uses CMC at rank-1, rank-5, and rank-10, plus mAP, on three benchmarks.
- Experimental setup: The search uses a ResNet macro structure with a 3x3 convolutional head and four sequential blocks containing neural cells.The search training and validation sets are formed by randomly splitting the official training images into two halves.
- Ablation studies: More than 2% mAP separates Auto-ReID from DARTS+PCB, GDAS+PCB, and Baseline+PCB in the part-module ablation.Auto-ReID searches both the number and location of the part-aware modules rather than attaching the part module only at the network tail.
- Benchmark comparisons: 94.5% rank-1 accuracy and 85.1% mAP are achieved on Market-1501, while reducing ResNet-50-based reID parameters by more than 45%.These results use the CNN found by Auto-ReID and exceed the reported baseline-search results of 93.8% rank-1 and 83.4% mAP.
- Benchmark comparisons: Auto-ReID significantly outperforms other models on both manually labeled and automatically detected CUHK03 bounding boxes.The CUHK03 comparison follows the new evaluation protocol.
- Benchmark comparisons: On MSMT17, Auto-ReID exceeds PCB by 12% mAP and 10% rank-1 accuracy.Across three benchmarks, the model outperforms most state-of-the-art methods and the baseline search algorithm, despite being searched on Market-1501.
- Limitations and future work: The proposed search space is limited to one reID-specific module, while the search algorithm is described as a simple extension of existing NAS.Future work proposes additional modules and more reID-specific search knowledge.
- Qualitative analysis: Automatically discovered cells are complex and difficult for human experts to find through manual tuning.The paper visualizes both normal and reduction cells from one searched architecture.
5. Conclusion
Auto-ReID automates neural architecture search for person re-identification by combining a body-structure-aware search space with retrieval-based searching. The resulting architecture significantly outperforms state-of-the-art reID models on three benchmarks.
- Auto-ReID combines a typical classification search space with a novel part-aware module that incorporates body structure information.The part-aware module is included among the candidate CNN components.
- Auto-ReID equips NAS with a retrieval loss to better match the reID task.The search algorithm is adapted because reID is fundamentally a retrieval task rather than classification.
- The architecture found by Auto-ReID significantly outperforms all state-of-the-art reID models on three benchmarks.