MFBN: An Efficient Base Model For Person Re-Identification

Person Re-IDentification (Re-ID) has developed rapidly with deep learning methods, as for these methods, the base mod- els used in most of them are not customized for Re-ID task. Although some studies have carefully designed the models special for Re-ID task, these models are always not easy to be the base model and expand with new methods due to their great complexity. In this paper, we propose a novel efficient base model named as Multi-granularity Feature Boosting Network (MFBN). MFBN consists of branches with information in different granularities. MFBN combines these branches into one whole, so MFBN is easy to be ex- tended as a base model. Moreover, MFBN applies feature boosting technique to boost fine granularitiy branch features with coarse granularity branch features, and applies channel- wise attention to increase diversities between features in multiple granularities. With these improves, MFBN has surpassed popular base models and got state-of-the-art results on mainstream Re-ID datasets including Market-1501, DukeMTMC-reID and CUHK-03. MFBN achieves results of rank-1/mAP=95.2%/93.2% on Market-1501 dataset and rank-1/mAP=90.3%/88.3% on DukeMTMC-reID dataset. Code and pretrained models are available at https://github.com/hsyi/Multi-granularity-Feature-Boosting-Network