Person Re-identification in Images and Videos
MetadataShow full item record
Person re-identification is a challenging task of matching a query person across multiple person's images or videos captured from different camera views. Recently, deep learning based approaches have showed promising performance on this task. In this thesis, initially we propose an image based person re-identification approach with Spatial Transformer Networks. Most previous deep learning based approaches use whole image features to compute the similarity between images. This is not very intuitive since not all the regions in an image contain information about the person identity. Hence, we introduce an end-to-end Siamese convolutional neural network that firstly localizes discriminative salient image regions and then computes the similarity based on these image regions. Furthermore, we propose an efficient attention based model for person re-identifying from videos. Our method generates an attention score for each frame based on frame-level features. The attention scores of all frames in a video are used to produce a weighted feature vector for the input video which is refined iteratively for re-identifying persons from videos. Extensive experiments on different datasets show that the proposed models provide an effective way of re-identifying person from images as well as videos.