Abstract: Deep learning-based models have achieved impressive performance on public segmentation benchmarks, yet generalizing to unseen environments remains challenging. Test-time training (TTT) ...