Adaptive Generalization for Semantic Segmentation
Out-of-distribution robustness remains a salient weakness of current state-of-the-art models for semantic segmentation. Until recently, research on generalization followed a restrictive assumption that the model parameters remain fixed after the training process. In this work, we empirically study an adaptive inference strategy for semantic segmentation that adjusts the model to the test sample before producing the final prediction. We achieve this with two complementary techniques. Using Instance-adaptive Batch Normalization (IaBN), we modify normalization layers by combining the feature statistics acquired at training time with those of the test sample. We next introduce a test-time training (TTT) approach for semantic segmentation, Seg-TTT, which adapts the model parameters to the test sample using a self-supervised loss. Relying on a more rigorous evaluation protocol compared to previous work on generalization in semantic segmentation, our study shows that these techniques consistently and significantly outperform the baseline and attain a new state of the art, substantially improving in accuracy over previous generalization methods.
PDF Abstract