A Self-Supervised Few-Shot Semantic Segmentation Method Based on Multi-Task Learning and Dense Attention Computation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Intelligent vehicles rely on vision systems for scene understanding, utilizing semantic segmentation.
• Traditional supervised semantic segmentation requires extensive pixel-level manual annotations, which is labor-intensive.
• Few-shot methods reduce annotation but remain demanding.

Purpose of the Study:

• To propose a novel self-supervised few-shot semantic segmentation method (MLDAC).
• To reduce the reliance on manual annotations in training semantic segmentation models for autonomous driving.
• To enhance the generalization ability and efficiency of perception systems.

Main Methods:

• Developed a self-supervised few-shot semantic segmentation approach using Multi-task Learning and Dense Attention Computation (MLDAC).
• Employed Swin Transformer as a backbone for multi-scale feature extraction.
• Integrated dense attention computation blocks and inter-scale mixing with feature skip connections.

Main Results:

• Achieved 55.1% and 26.8% one-shot mIoU on PASCAL-5i and COCO-20i datasets, respectively.
• Demonstrated strong performance with 78.1% accuracy on the FSS-1000 dataset.
• Validated the efficacy of the proposed MLDAC method in self-supervised few-shot segmentation tasks.

Conclusions:

• MLDAC effectively reduces annotation burden for semantic segmentation in autonomous driving.
• The method shows significant improvements in few-shot segmentation accuracy and generalization.
• MLDAC offers a promising direction for efficient scene understanding in intelligent vehicles.