Unsupervised Object Re-identification via Irregular Sampling

Object re-identification (Re-ID), is a fundamental task in intelligent systems, that aims to find the same object, i.e., person or vehicle under different camera views or scenes. This paper studies the fully unsupervised object re- ID problem which can learn re- ID without any human-annotated labeled data. Recent works show that self-supervised momentum contrastive learning is an effective method for unsupervised object re- ID, but they neglect to optimize one important component - sampling strategy. Here we investigate and analyze the performances of the current sampling strategy in different numbers of positive samples in a mini-batch under the same learning framework and loss function, then we proposed a more effective and robust sampling strategy - Irregular Sampling (IS). Experimental results show that sampling strategy is also an important factor in model performance, and the proposed sampling strategy IS can effectively boost the model performance. Extensive experiments are performed on one vehicle re-ID dataset and two mainstream person re- ID datasets.