Pseudolabel guided pixels contrast for domain adaptive semantic segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Semantic segmentation requires pixel-level annotations, which are expensive and time-consuming to acquire.
• Unsupervised Domain Adaptation (UDA) for semantic segmentation leverages labeled synthetic data for unlabeled real-world data.
• Existing UDA methods using contrastive learning overlook intra-class feature diversity, causing prediction errors.

Purpose of the Study:

• To address the limitations of current contrastive learning approaches in UDA for semantic segmentation.
• To propose a novel framework, Pseudo-label Guided Pixel Contrast (PGPC), to improve class prediction accuracy.
• To effectively utilize information from target images while minimizing pseudo-label noise.

Main Methods:

• Developed the Pseudo-label Guided Pixel Contrast (PGPC) framework.
• Incorporated contrastive learning strategies that consider intra-class feature diversity.
• Investigated methods to leverage target image information without introducing noise.

Main Results:

• PGPC framework demonstrated superior performance on standard UDA benchmarks.
• Achieved relative improvements of 5.1% mIoU (GTA5 to Cityscapes) and 4.6% mIoU (SYNTHIA to Cityscapes) using DAFormer.
• The proposed approach enhances existing UDA methods without increasing model complexity.

Conclusions:

• PGPC effectively overcomes limitations in previous UDA methods for semantic segmentation.
• The framework improves model accuracy by considering feature diversity within classes.
• PGPC offers a valuable enhancement for various UDA approaches in semantic segmentation tasks.