2020/11/25 by Marco Toldo, Toldo, Marco, Umberto Michieli +3
Computer Science · #Domain Adaptation and Few-Shot Learning #Multimodal Machine Learning Applications
paper · pdf · doi:10.48550/arxiv.2011.12616
Deep learning frameworks allowed for a remarkable advancement in semantic\nsegmentation, but the data hungry nature of convolutional networks has rapidly\nraised the demand for adaptation techniques able to transfer learned knowledge\nfrom label-abundant domains to unlabeled ones. In this paper we propose an\neffective Unsupervised Domain Adaptation (UDA) strategy, based on a feature\nclustering method that captures the different semantic modes of the feature\ndistribution and groups features of the same class into tight and\nwell-separated clusters. Furthermore, we introduce two novel learning\nobjectives to enhance the discriminative clustering performance: an\northogonality loss forces spaced out individual representations to be\northogonal, while a sparsity loss reduces class-wise the number of active\nfeature channels. The joint effect of these modules is to regularize the\nstructure of the feature space. Extensive evaluations in the synthetic-to-real\nscenario show that we achieve state-of-the-art performance.\n