Learning Rubost Features with Refined Spatial Interactions of Inter-part for Person Re-IDentification

The combination of the part features and attention mechanism has been made remarkable progress in Person Re-IDentification (REID), but part-based approaches are still ignoring the relations of inter-part and missing discriminative local feature representations. The Refined Spatial Interactions of Inter-part for Person Re-IDentification (RSIPR) is proposed to solve above problems. (RSIPR) consists of two main components: 1) Self Filter Attention (SFA) mechanism is employed to extract discriminatory representations from the pedestrian image adaptively. 2) Refined Spatial Interactions (RSI) is utilized to refine the interactions of inter-part and avoid ignoring the global perception due to the pedestrain is cropped into patches directly. Extensive experiments have been implemented to validate the superiority of the (RSIPR), and it has been demonstrated that the (RSIPR) achieves the state-of-the-art method without re-ranking [1] algorithm.