AdaTriplet-RA: Domain matching via adaptive triplet and reinforced attention for unsupervised domain adaptation

Unsupervised domain adaptation (UDA) is a transfer learning task where the annotations of source are available, but only have access to unlabeled target data during training. Previous methods minimise gap by performing distribution alignment between and domains, which has notable limitation, i.e., at level, neglecting sample-level differences, thus preventing model from achieving superior performance. To solve this, we improve UDA with an inter-domain matching scheme. We apply widely-used Triplet loss match samples. reduce catastrophic effect inaccurate pseudo-labels generated training, propose novel uncertainty measurement method use this select reliable automatically. As selection uncertainty-aware, pseudo labels progressively refined as training performed. advanced Gumbel Softmax technique realise adaptive Top-k scheme achieve selection. enable global ranking optimisation within one batch for matching, whole optimised via reinforced attention mechanism, using Average Precision (AP) reward. Our AdaTriplet-RA achieves State-of-the-art results on several public benchmark datasets, its effectiveness validated comprehensive ablation study. improves accuracy baseline 9.7% (using ResNet-101 backbone network) 6.2% (ResNet-50) VisDa dataset 4.22% DomainNet dataset. The code publicly available at: https://github.com/shuxy0120/AdaTriplet-RA.