Video person re-identification using key frame screening with index and feature reorganization based on inter-frame relation

Nowadays, there are many video person re-identification networks that do not consider screening the input video frame sequence, which result in the high-similarity of the video frames used for training the neural network. In this way, the temporal information in the video cannot be effectively modeled. To address that, we try to propose a video person re-identification scheme based on inter-frame reorganization, which consists of two modules. First, the Key Frame Screening with Index (KFSI) is proposed to screen the similar frames, and a frame sequence with richer information is extracted when loading the training dataset. Second, the Feature Reorganization Based on Inter-Frame Relation (FRBIFR) is proposed to reorganize the features of key frame sequence by calculating the correlation between the frames, and the reorganized features are more robust by eliminating some distractions (such as occlusion etc.). The experimental results show that our method outperforms the state-of-the-art methods on four mainstream datasets MARS, ILIDS-VID, PRID-2011 and DukeMTMC-VideoReID.