Large-Scale Video-Based Person Re-identification via Non-local Attention and Feature Erasing

Abstract

Encoding the video tracks of person to an aggregative representation is the key for video-based person re-identification (re-ID), where average pooling or RNN methods are typically used to aggregating frame-level features. However, It is still difficult to deal with the spatial misalignment caused by occlusion, posture changes and camera views. Inspired by the success of non-local block in video analysis, we use a non-local attention block as a spatial-temporal attention mechanism to handle the spatial-temporal misalignment problem. Moreover, partial occlusion is widely occurred in video sequences. We propose a local feature branch to tackle the partial occlusion problem by using feature erasing in the frame-level feature map. Therefore, our network is composed by two-branch, the global branch via non-local attention encoding the global feature and the local feature branch grasping the local feature. In evaluation, the global feature and local feature are concatenated to obtain a more discriminative feature. We conduct extensive experiments on two challenging datasets (MARS and iLIDS-VID). The experimental results demonstrate that our method is comparable with the state-of-the-art methods in these datasets.