Learning Instance-Specific Adaptation for Cross-Domain Segmentation

We propose a test-time adaptation method for cross-domain image segmentation. Our is simple: Given new unseen instance at test time, we adapt pre-trained model by conducting instance-specific BatchNorm (statistics) calibration. approach has two core components. First, replace the manually designed calibration rule with learnable module. Second, leverage strong data augmentation to simulate random domain shifts learning rule. In contrast existing methods, our does not require accessing target training time or computationally expensive training/optimization. Equipping models trained standard recipes achieves significant improvement, comparing favorably several state-of-the-art generalization and one-shot unsupervised approaches. Combining methods further improves performance, reaching state of art. project page https://yuliang.vision/InstCal/ .