Subsidiary Prototype Alignment for Universal Domain Adaptation

• DOI: 10.48550/arxiv.2210.15909
• Publication Date: 2022-01-01
• Authors: Jogendra Nath Kundu; Suvaansh Bhambri; Akshay Kulkarni; Hiran Sarkar; Varun Jampani; R. Venkatesh Babu
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition; 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology; Machine Learning (cs.LG)

Universal Domain Adaptation (UniDA) deals with the problem of knowledge transfer between two datasets domain-shift as well category-shift. The goal is to categorize unlabeled target samples, either into one "known" categories or a single "unknown" category. A major in UniDA negative transfer, i.e. misalignment and classes. To this end, we first uncover an intriguing tradeoff negative-transfer-risk domain-invariance exhibited at different layers deep network. It turns out can strike balance these metrics mid-level layer. Towards designing effective framework based on insight, draw motivation from Bag-of-visual-Words (BoW). Word-prototypes BoW-like representation layer would represent lower-level visual primitives that are likely be unaffected by category-shift high-level features. We develop modifications encourage learning word-prototypes followed word-histogram classification. Following this, subsidiary prototype-space alignment (SPA) seen closed-set problem, thereby avoiding transfer. realize novel word-histogram-related pretext task enable SPA, operating conjunction UniDA. demonstrate efficacy our approach top existing techniques, yielding state-of-the-art performance across three standard Open-Set DA object recognition benchmarks.