Collaborative Attention Network for Person Re-identification

The quality of visual feature representation has always been a key factor in many computer vision tasks. In the person re-identification (Re-ID) problem, combining global and local features to improve model performance is becoming a popular method, because previous works only used global features alone, which is very limited at extracting discriminative local patterns from the obtained representation. Some existing works try to collect local patterns explicitly slice the global feature into several local pieces in a handcrafted way. By adopting the slicing and duplication operation, models can achieve relatively higher accuracy but we argue that it still does not take full advantage of partial patterns because the rule and strategy local slices are defined. In this paper, we show that by firstly over-segmenting the global region by the proposed multi-branch structure, and then by learning to combine local features from neighbourhood regions using the proposed Collaborative Attention Network (CAN), the final feature representation for Re-ID can be further improved. The experiment results on several widely-used public datasets prove that our method outperforms many existing state-of-the-art methods.