Towards Unsupervised Domain Bridging via Image Degradation in Semantic Segmentation

TL;DR

Domain shift remains a core challenge for semantic segmentation under unsupervised domain adaptation. The paper introduces DiDA, a degradation-based domain bridging framework that uses a forward diffusion process to create intermediate degraded domains and a diffusion encoder with time conditioning to compensate semantic shift, reinforced by a degraded-image consistency loss. The approach is plug-and-play, compatible with existing UDA methods, and extensible to arbitrary degradations, demonstrated by consistent performance gains across multiple benchmarks and backbones, including a new state-of-the-art on MIC. By linking degradation overlap to a domain-shared prior, DiDA provides a practical and general strategy for learning domain-invariant features in segmentation tasks.

Abstract

Semantic segmentation suffers from significant performance degradation when the trained network is applied to a different domain. To address this issue, unsupervised domain adaptation (UDA) has been extensively studied. Despite the effectiveness of selftraining techniques in UDA, they still overlook the explicit modeling of domain-shared feature extraction. In this paper, we propose DiDA, an unsupervised domain bridging approach for semantic segmentation. DiDA consists of two key modules: (1) Degradation-based Intermediate Domain Construction, which creates continuous intermediate domains through simple image degradation operations to encourage learning domain-invariant features as domain differences gradually diminish; (2) Semantic Shift Compensation, which leverages a diffusion encoder to disentangle and compensate for semantic shift information with degraded timesteps, preserving discriminative representations in the intermediate domains. As a plug-and-play solution, DiDA supports various degradation operations and seamlessly integrates with existing UDA methods. Extensive experiments on multiple domain adaptive semantic segmentation benchmarks demonstrate that DiDA consistently achieves significant performance improvements across all settings. Code is available at https://github.com/Woof6/DiDA.

Figures (17)

• Figure 1: Conceptual illustration of the diffusion forward process. Fine-grained, domain-specific attributes such as texture are lost with less noise added (i.e., early time-steps), while coarse-grained, domain-invariant ones such as shape are lost by adding more noise (i.e., late time-steps).
• Figure 1: Quantitative results of DiDA on different methods and benchmarks with CNN-based model (C) or Transformer-based model (T). $*$ denotes the reproduced result.
• Figure 2: Overview of DiDA framework. We integrate diffusion strategies (green box) with a standard self-training paradigm. While regular frameworks train networks using supervised loss on source domain and unsupervised adaptation loss on target domain, DiDA introduces degradation-based intermediate domains and addresses semantic shift through a diffusion encoder and reconstruction head, which are enabled by degraded image consistency (DIC) loss and reconstruction loss.
• Figure 3: Demonstration of two modes of inference.
• Figure 4: The performance variation with the degraded level.