SCMix: Stochastic Compound Mixing for Open Compound Domain Adaptation in Semantic Segmentation

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Open Compound Domain Adaptation (OCDA) transfers knowledge from labeled source domains to unlabeled target domains, including unseen ones.
• Current OCDA methods use a divide-and-conquer approach, potentially underestimating target domain variance.
• This can limit model generalization and performance in complex adaptation scenarios.

Purpose of the Study:

• To establish a novel generalization bound for OCDA based on general domain adaptation theory.
• To propose a new augmentation strategy, Stochastic Compound Mixing (SCMix), to mitigate distribution divergence between source and target domains.
• To theoretically and empirically demonstrate the superiority of SCMix over conventional OCDA techniques.

Main Methods:

• Developed a novel generalization bound specifically for the OCDA setting.
• Introduced Stochastic Compound Mixing (SCMix), an augmentation strategy designed to reduce the divergence between source and mixed target distributions.
• Conducted theoretical analyses to prove SCMix's advantages, showing single-target mixing as a subset of the proposed method.

Main Results:

• Theoretical analyses confirmed the effectiveness of SCMix, with single-target mixing identified as a subgroup.
• Extensive experiments on OCDA semantic segmentation tasks demonstrated lower empirical risk using SCMix.
• Combining SCMix with transformer architectures yielded significant performance improvements over state-of-the-art results.

Conclusions:

• Conventional OCDA methods may underestimate target domain variance, hindering generalization.
• SCMix effectively mitigates distribution divergence, leading to improved model performance in OCDA.
• The proposed method, particularly when combined with transformers, represents a significant advancement in OCDA for semantic segmentation.