Unsupervised Person Re-Identification with Iterative Self-Supervised Domain Adaptation

Abstract

In real applications, person re-identification (re-id) is an inherently domain adaptive computer vision task which often requires the model trained on a group of people to perform well on an unlabeled dataset consisting of another group of pedestrians without supervised fine-tuning. Furthermore, there are typically a large number of classes (people) with small number of samples belonging to each class. Based on the characteristics of person re-id and general assumptions related to domain adaptation, we put forward a novel algorithm for cross-dataset person re-id. Our idea is simple yet effective: first, we preprocess the source dataset with style transfer GAN and train a baseline on it in a supervised learning manner, then we assign pseudo labels to unlabeled samples in target dataset based on the model trained on labeled source dataset; finally, we train on the target dataset with pseudo labels in traditional supervised learning manner. We adopt the idea of co-training in the training process to make the pseudo labels more reliable. We show the superiority of our model over all state-of-the-art methods through extensive experiments.