2016/11/24 by Piotr Koniusz, Koniusz, Piotr, Yusuf Tas +3 · 2 citations
Computer Science · #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Speech and Audio Processing #Video Surveillance and Tracking Methods
paper · pdf · doi:10.48550/arxiv.1611.08195
openalex publication_date 2016/11/24 · openalex created_date 2022/10/03 · openalex updated_date 2026/07/28
In this paper, we propose an approach to the domain adaptation, dubbed\nSecond- or Higher-order Transfer of Knowledge (So-HoT), based on the mixture of\nalignments of second- or higher-order scatter statistics between the source and\ntarget domains. The human ability to learn from few labeled samples is a\nrecurring motivation in the literature for domain adaptation. Towards this end,\nwe investigate the supervised target scenario for which few labeled target\ntraining samples per category exist. Specifically, we utilize two CNN streams:\nthe source and target networks fused at the classifier level. Features from the\nfully connected layers fc7 of each network are used to compute second- or even\nhigher-order scatter tensors; one per network stream per class. As the source\nand target distributions are somewhat different despite being related, we align\nthe scatters of the two network streams of the same class (within-class\nscatters) to a desired degree with our bespoke loss while maintaining good\nseparation of the between-class scatters. We train the entire network in\nend-to-end fashion. We provide evaluations on the standard Office benchmark\n(visual domains), RGB-D combined with Caltech256 (depth-to-rgb transfer) and\nPascal VOC2007 combined with the TU Berlin dataset (image-to-sketch transfer).\nWe attain state-of-the-art results.\n