Efficient Person re-identification with Multi-Scale Feature Fusion

The task of Person Re-Identification (Re-ID) has attracted growing attention in recent years. Most State-Of-The-Art (SOTA) methods often employ the large-scale ResNet model as their backbone, which makes it tedious to explore various architecture modifications. In this study, we propose a small-sized Re-ID model with EfficientNet as its backbone, along with a novel feature pyramid branch for learning multi-scale features. Extensive experimental results show that the proposed network outperforms various SOTA methods with obvious margin on standard benchmark datasets such as Market1501, DukeMTMC-Re-ID, CUHK03, even though its mode only contains about 7.41M parameters.