Where to look: Multi-granularity occlusion aware for video person re-identification

Abstract

Video person re-identification(re-ID) plays an important role in intelligent video surveillance, which can automatically match the same person across video clips under non-overlapping cameras. Despite great progress in re-ID, the performance of most existing methods still is corrupted severely under partial occlusion. Although some multi-granularity methods have alleviated this dilemma, these methods still suffer from weak diversity of features and conflict between rigid horizontal partition and vertical occlusion. In this paper, we propose a novel video person re-ID framework, called Multi-Granularity Occlusion Aware (MGOA), which extracts multi-granularity features by precisely erasing the occlusion. Different from previous works based on multiple granularities, the proposed MGOA predicts the partial occlusion in a coarse-to-fine manner instead of erasing the occlusion in video clips by one step. Specifically, we first propose the multi-granularity feature extraction to obtain diverse features at different levels of feature maps, which is beneficial for the fine erasure of the occlusion. Moreover, to avoid the limitation of horizontal stripes that cannot handle vertical occlusion, we design Attention-Aware Occlusion Erasure (AA-OE) that can obtain the attention maps with coarse occlusion erasure in the coarse-grained branch and the attention maps with fine occlusion erasure in the fine-grained branch. It is worth noting that each granularity in our network is not independent but relevant through the top-down information transmission between granularities, which transfers the erased occlusion feature maps of the current branch to the next finer-grained branch for guiding AA-OE to obtain more discriminative features. Extensive experiments on three challenging public benchmarks show that our MGOA can deal well with occlusion and achieves state-of-the-art performance.