Masked adaptive diffusion for unsupervised domain-adaptive semantic segmentation

Robots and other autonomous systems often rely on semantic segmentation for scene understanding, but segmentation accuracy degrades under deployment-time domain shifts, such as weather and illumination changes or the synthetic-to-real gap. Unsupervised domain adaptation (UDA) mitigates this by transferring knowledge from labeled source data to unlabeled target data. Diffusion models can progressively transform distributions, but directly coupling diffusion-based feature translation with self-training UDA can produce a highly variable directional relationship between the segmentation and auxiliary gradients along the realized training trajectory, including disagreement between the two objectives. To improve this coupling, we propose Masked Adaptive Diffusion (MAD), an optimization-aware diffusion framework. MAD first constructs domain-bridging representations by applying forward diffusion to source features and then performing truncated reverse diffusion, so that the diffusion timestep controls the degree of source-to-target translation. It then introduces an SNR-guided adaptive masking strategy designed to attenuate unreliable feature perturbations within a single forward pass and promote more consistent directional compatibility. Because the diffusion and masking modules are training-only, MAD leaves the deployed segmentation network and its onboard inference cost unchanged. Experiments on synthetic-to-real GTA→Cityscapes and clear-to-adverse-weather Cityscapes→ACDC driving benchmarks show that MAD improves strong self-training baselines and achieves 77.4 and 72.6 mIoU, respectively.