Unsupervised visual feature learning based on similarity guidance

The availability of a large amount of image data and the impracticality of annotating each sample, coupled with various changes in the target class, such as lighting, posture, etc., make the performance of feature learning disappointing on unlabeled datasets. Lack of attention to hard sample pairs in network modeling and one-sided consideration of similarity measurement in the process of merging have exacerbated the huge performance gap between supervised and unsupervised feature expression. In order to alleviate these problems, we propose an unsupervised network that gradually optimizes feature expression under the guidance of similarity. It employs the deep network to train high-dimensional features and small-scale merge to generate high-quality labels to alternately execute the two steps. Feature learning is guided by gradually generating high-quality labels, thereby narrowing the huge gap between unsupervised learning and supervised learning. The proposed method has been evaluated on both general datasets and the datasets for person re-identification (person re-ID) with superior performance in comparison with existing state-of-the-art methods.