2020/06/25 by Yue Cao, Cao, Yue, Zhenda Xie +9 · 18 citations
Computer Science · #Advanced Image and Video Retrieval Techniques #Advanced Vision and Imaging #Artificial intelligence #Computer Vision and Pattern Recognition (cs.CV) #Computer science #FOS: Computer and information sciences #Feature (linguistics) #Image Retrieval and Classification Techniques #Machine Learning (cs.LG) #Machine learning #Parametric statistics #Pattern recognition (psychology) #Simple (philosophy) #Sliding window protocol #Window (computing) #cs.CV #cs.LG
paper · pdf · doi:10.48550/arxiv.2006.14618
published in arXiv (Cornell University) (Cornell University)
arxiv created 2020/06/25 · openalex publication_date 2020/06/25 · arxiv updated 2020/06/26 · openalex created_date 2020/11/23 · openalex updated_date 2026/08/05
This paper presents parametric instance classification (PIC) for unsupervised\nvisual feature learning. Unlike the state-of-the-art approaches which do\ninstance discrimination in a dual-branch non-parametric fashion, PIC directly\nperforms a one-branch parametric instance classification, revealing a simple\nframework similar to supervised classification and without the need to address\nthe information leakage issue. We show that the simple PIC framework can be as\neffective as the state-of-the-art approaches, i.e. SimCLR and MoCo v2, by\nadapting several common component settings used in the state-of-the-art\napproaches. We also propose two novel techniques to further improve\neffectiveness and practicality of PIC: 1) a sliding-window data scheduler,\ninstead of the previous epoch-based data scheduler, which addresses the\nextremely infrequent instance visiting issue in PIC and improves the\neffectiveness; 2) a negative sampling and weight update correction approach to\nreduce the training time and GPU memory consumption, which also enables\napplication of PIC to almost unlimited training images. We hope that the PIC\nframework can serve as a simple baseline to facilitate future study.\n