Cross-domain Person Re-identification

Abstract

The problem of person reidentification has been getting much attention in the computer vision community. The task is to recognize pictures of the same indi� viduals in images with different backgrounds taken from multiple cameras. It involves complexities such as different people with similar outfits or the same person with different outfits, differential illuminations, lowresolution images, in� accurate bounding boxes, and occlusion. The increasing progress is due to the increased demand for automated surveillance systems. The problem is often formulated as a retrieval task. When given a query image and a gallery set of images taken from different cameras, possibly at different locations, the system aims to find the pictures with the same person. In this thesis, we have used ResNetSO and Omni-scale Network (OSNet) for feature extraction and different loss functions such as softmax loss, triplet loss, quadruplet loss, KL-divergence, etc. to train and infer on models for cross-domain person re-identification (re-id). We observe that using a multi-source training strategy boosts the performance of such cross-domain re-id systems. We also show that using re-ranking significantly improves the performance of both same domain and cross-domain person re-id.