57 / 2022-09-29 14:35:07

Searching a Multi-level Aggregation Architecture for Person Re-Identification

Modeling both global information and local details is a trend for high performance person re-identification (ReID). However, most of current works manually design feature extraction frameworks, which require highly massive prior knowledge, ~\emph{e.g.,} experience and trial. To solve this issue, we propose a novel approach to automatically find an effective architecture based Neural Architecture Search (NAS), aiming at imitating the process of human visual perception to filter background noises  and concentrate on human bodies and local identity-related information. Specifically, we construct a multi-level aggregation architecture to capture and aggregate more representative features from different levels. Then, we adopt a task-specific search space, including two search cells and effective operators. Finally, an optimized architecture can be obtained by an efficient cooperative search strategy that is introduced to explore the search space from cell-level and operation-level. Equipped with the searched architectures, extensive experiments verify that our method achieves state-of-the-art performance on four person ReID benchmarks.