PMT-Net: Progressive Multi-Task Network for one-shot Person Re-Identification

Abstract

Most existing person Re-Identification (Re-ID) algorithms require abundant labeled data from paired non-overlapping camera views in the fully supervised scenario. However, the fully supervised Re-ID suffers from the limited availability of labeled training samples due to the sharply increased cost of manual efforts. To tackle this problem, a novel Progressive Multi-Task Network (PMT-Net) for person Re-ID is proposed. PMT-Net initializes a model using only one labeled sample for each identity, and it iteratively optimizes the model by sampling the most reliable pseudo labels dynamically from unlabeled samples. Firstly, pedestrian attributes recognition is incorporated as an auxiliary task to learn discriminative features. Then, based on the discriminative features, the identity label for unlabeled samples is estimated by the distance between the labeled samples and unlabeled samples in feature space. In addition, to enhance the accuracy of label estimation for the unlabeled samples, a semi-supervised clustering method, named Distance Ranked Weight Clustering (DRW-Clustering) is designed. The clustering method weights partial unlabeled samples by the indexed ordinal of distance sorting, so that it can find the real cluster center quickly and effectively. Extensive comparative evaluation experiments are conducted on Market1501 and DukeMTMC-reID datasets, and the experimental results indicate that the proposed method achieves performance competitive or better than that of the state-of-the-art for one-shot person Re-ID.