Hard-Aware Instance Adaptive Self-Training for Unsupervised Cross-Domain Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning models face challenges with data divergence between training and testing sets.
• Unsupervised Domain Adaptation (UDA) addresses this by adapting models to new data distributions without labeled target data.
• Self-training is a promising UDA technique, but existing methods struggle with scalability and performance trade-offs.

Purpose of the Study:

• To propose a novel hard-aware instance adaptive self-training framework for UDA in semantic segmentation.
• To enhance the quality and diversity of pseudo-labels generated during the self-training process.
• To improve the performance and generalization capabilities of UDA models.

Main Methods:

• Developed a pseudo-label generation strategy with an instance adaptive selector.
• Introduced hard-aware pseudo-label augmentation using inter-image information for challenging classes.
• Implemented region-adaptive regularization for pseudo-label smoothing and non-pseudo-label sharpening.
• Incorporated consistency constraints for non-pseudo-label regions to strengthen supervision.

Main Results:

• The proposed framework demonstrated superior performance compared to state-of-the-art methods.
• Experiments were conducted on challenging cross-domain datasets: GTA5 $\rightarrow$ Cityscapes, SYNTHIA $\rightarrow$ Cityscapes, and Cityscapes $\rightarrow$ Oxford RobotCar.
• The method effectively improved pseudo-label quality and model adaptation.

Conclusions:

• The hard-aware instance adaptive self-training framework offers a concise and efficient solution for UDA in semantic segmentation.
• The proposed techniques for pseudo-label generation and augmentation significantly boost model performance.
• The framework is generalizable and shows strong potential for various UDA applications.