Exploring Category Attention for Open Set Domain Adaptation

Great success has been achieved in the area of unsupervised domain adaptation, which learns to generalize from labeled source unlabeled target domain. However, most existing techniques can only handle closed-set scenario, requires both and have a shared category label set. In this work, we propose two-stage method deal with more challenging task open set where contains categories unseen Our first stage formulates alignment two domains as semi-supervised clustering problem, initially associates each target-domain sample x t ∈ X source-domain ℓ xmlns:xlink="http://www.w3.org/1999/xlink">s L . To end, use self-training strategy learn teacher network student network, adopt self-attention mechanism. second refines resulting clusters by identifying negative associations (x , ) labeling involved xt unknown. For purpose, investigate compatibility association replacing maps last convolutional layers newly proposed attention (CAMs), locate informative feature pixels for given category. Experimental results on three public datasets show effectiveness robustness our adaptation across various pairs.