Unsupervised Model Adaptation for Continual Semantic Segmentation

We develop an algorithm for adapting a semantic segmentation model that is trained using labeled source domain to generalize well in unlabeled target domain. A similar problem has been studied extensively the unsupervised adaptation (UDA) literature, but existing UDA algorithms require access both data and training agnostic model. Relaxing this constraint enables user adapt pretrained models domain, without requiring data. To end, we learn prototypical distribution intermediate embedding space. This encodes abstract knowledge learned from then use aligning with provide theoretical analysis explain conditions under which our effective. Experiments on benchmark tasks demonstrate method achieves competitive performance even compared joint approaches.