Feature Space Regularization for Person Re-identification with One Sample

Abstract

Few Shot Learning is a solution to relieve the huge annotation cost in Person Re-Identification. We concentrate on one sample setting in this work, where each identity has only one labeled sample along with many unlabeled samples. Training with one sample setting, the model is easily biased towards certain identities. Moreover, a reliable pseudo-label estimation scheme can greatly improve the final performance of the model. Targeting to solve the issues above, we propose two simple and effective solutions. (a) We design the Feature Space Regularization (FSR) Loss to adjust the distribution of samples in feature space. The FSR loss make the difference in distance of all labeled samples to unlabeled samples as small as possible. (b) We propose combining the Nearest Neighbor distance with inter-class distance to estimate pseudo-label for unlabeled data, which we called Joint-Distance. Notably, the Rank-1 accuracy of our method outperforms the state of the art method by a large margin of 12.1 points (absolute, i.e., 67.9% vs. 55.8%) on Market-1501, and 10.1 points (absolute, i.e., 58.9% vs. 48.8%) on DukeMTMC-reID, respectively. We will release all the code in https://github.com/Freedomxt/Feature_Space_Regularization_for_person_Re-Identification_with_One_Sample.