Semantic Segmentation Based Unsupervised Domain Adaptation via Pseudo-Label Fusion

In this paper, we propose a pseudo label fusion framework (PLF), a learning framework developed to deal with the domain gap between a source domain and a target domain for performing semantic segmentation based UDA in the unseen target domain. PLF fuses the pseudo labels generated by an ensemble of teacher models. The fused pseudo labels are then used by a student model to distill out the information embedded in these fused pseudo labels to perform semantic segmentation in the target domain. To examine the effectiveness of PLF, we perform a number of experiments on both GTA5 to Cityscapes and SYNTHIA to Cityscapes benchmarks to quantitatively and qualitatively inspect the improvements achieved by employing PLF in performing semantic segmentation in the target domain. Moreover, we provide a number of parameter analyses to validate that the choices made in the design of PLF is both practical and beneficial. Our experimental results on both benchmarks shows that PLF indeed offers adequate performance benefits in performing semantic segmentation in the unseen domain, and is able to achieve competitive performance when compared to the contemporary UDA techniques.