SePiCo: Semantic-Guided Pixel Contrast for Domain Adaptive Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Domain adaptive semantic segmentation aims to predict pixel-level labels for unlabeled target data using models trained on labeled source data.
• Self-training with pseudo-labels is a common approach, but existing methods often overlook the importance of intra-class compactness and inter-class dispersion in pixel representations.
• This oversight leads to challenges in handling cross-domain semantic variations and results in poorly structured embedding spaces, hindering generalization.

Purpose of the Study:

• To propose Semantic-Guided Pixel Contrast (SePiCo), a novel one-stage domain adaptation framework for semantic segmentation.
• To enhance self-training methods by learning class-discriminative and class-balanced pixel representations across domains.
• To address the limitations of existing methods in handling noisy pseudo-labels and cross-domain variations.

Main Methods:

• SePiCo employs a centroid-aware pixel contrast to guide feature learning using category centroids from the source domain.
• It introduces a distribution-aware pixel contrast to capture a sufficient quantity of instances by approximating category distributions from source data statistics.
• The framework optimizes pixel representations for improved intra-class compactness and inter-class dispersion, leading to computationally efficient adaptation.

Main Results:

• SePiCo stabilizes the training process for domain adaptive semantic segmentation.
• The method yields highly discriminative pixel representations, significantly improving performance in both synthetic-to-real and daytime-to-nighttime adaptation scenarios.
• Experimental results demonstrate substantial progress in addressing cross-domain semantic variations and enhancing generalization.

Conclusions:

• SePiCo effectively addresses the limitations of previous domain adaptation methods by focusing on pixel-level semantic relationships.
• The proposed contrastive learning approach enhances the quality of pseudo-labels and the structure of the embedding space.
• SePiCo offers a promising direction for improving the robustness and accuracy of semantic segmentation in diverse, unlabeled target domains.