DA-Cal: Towards Cross-Domain Calibration in Semantic Segmentation

TL;DR

DA-Cal is proposed, a dedicated cross-domain calibration framework that transforms target domain calibration into soft pseudo-label optimization and seamlessly integrates with existing self-training frameworks across multiple UDA segmentation benchmarks, significantly improving target domain calibration while delivering performance gains without inference overhead.

Abstract

While existing unsupervised domain adaptation (UDA) methods greatly enhance target domain performance in semantic segmentation, they often neglect network calibration quality, resulting in misalignment between prediction confidence and actual accuracy -- a significant risk in safety-critical applications. Our key insight emerges from observing that performance degrades substantially when soft pseudo-labels replace hard pseudo-labels in cross-domain scenarios due to poor calibration, despite the theoretical equivalence of perfectly calibrated soft pseudo-labels to hard pseudo-labels. Based on this finding, we propose DA-Cal, a dedicated cross-domain calibration framework that transforms target domain calibration into soft pseudo-label optimization. DA-Cal introduces a Meta Temperature Network to generate pixel-level calibration parameters and employs bi-level optimization to establish the relationship between soft pseudo-labels and UDA supervision, while utilizing complementary domain-mixing strategies to prevent overfitting and reduce domain discrepancies. Experiments demonstrate that DA-Cal seamlessly integrates with existing self-training frameworks across multiple UDA segmentation benchmarks, significantly improving target domain calibration while delivering performance gains without inference overhead. The code will be released.

Figures (7)

• Figure 1: Reliability Diagrams niculescu2005predicting. We observe that neural networks not only degrade in performance under cross-domain settings but also exhibit poor calibration. Although existing UDA methods, such as DACS tranheden2021dacs and DAFormer hoyer2022daformer), significantly improve target-domain performance, they still struggle to produce well-calibrated predictions. When integrated with our proposed DA-Cal, these methods achieve consistently improved calibration, leading to more reliable confidence estimates.
• Figure 2: Pipeline illustration of DA-Cal. Our framework consists of bi-level optimization: inner optimization (Steps 1-2) for calibrating predictions with Meta Temperature Network, and outer optimization (Step 3) for training the segmentation network with both calibrated soft and hard pseudo-labels. The supervised loss on source domain is omitted for clarity.
• Figure 3: Comparison of DA-Cal and Baseline Methods on Reliability Diagrams.
• Figure 4: Ablation study on temperature scaling for GTAv. $\rightarrow$CS. benchmark across different methods.
• Figure 5: Visualization of temperature maps on different benchmarks.