Person re-identification with neural architecture search

Abstract

Most of the existing person re-identification (ReID) methods use a classification network pre-trained on external data as the backbone and then fine-tune it, which results in a network architecture that is fixed and dependent on pre-training of external data. There are also some methods that are specifically designed by human experts for ReID, but manual network design becomes more difficult as network requirements increase and often fails to achieve optimal settings. In this paper, we consider using emerging neural architecture search (NAS) technology as a tool to solve above problems. However, most of NAS methods deal with classification tasks, which causes NAS to not be directly extended to ReID. In order to coordinate the inconsistency between the two optimization goals, we propose to establish an objective function with the assistant of the triplet loss to guide the direction of architecture search. Finally, it is no longer dependent on external data to automatically generate a ReID network with excellent performance using NAS directly on the target dataset. The experimental results on three public datasets validate that our method can automatically and efficiently find the network architecture suitable for ReID.