Lidar-Camera Semi-Supervised Learning for Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Semantic segmentation is crucial for autonomous systems.
• Individual sensor modalities (LiDAR, camera) have limitations in diverse environmental conditions.
• Data annotation for supervised learning is costly and time-consuming.

Purpose of the Study:

• To evaluate the performance improvement of fused LiDAR and camera data for semantic segmentation.
• To investigate the application of semi-supervised learning with sensor fusion for domain adaptation.
• To assess the effectiveness of fusion and semi-supervised learning in challenging, real-world scenarios.

Main Methods:

• Comparative study of neural networks trained with individual sensors versus fused sensor data.
• Experimental evaluation across various scenarios including adverse weather and lighting conditions.
• Implementation of semi-supervised learning techniques to leverage unlabeled data.

Main Results:

• Sensor fusion significantly improves semantic segmentation performance compared to individual LiDAR or camera data.
• Semi-supervised learning combined with sensor fusion enhances model robustness in challenging scenarios.
• The proposed approach reduces the need for extensive labeled data, particularly for domain adaptation.

Conclusions:

• Fusion of LiDAR and camera data is highly effective for robust semantic segmentation.
• Semi-supervised learning offers a viable approach to adapt models to new domains with reduced annotation effort.
• The combined techniques provide superior performance in complex and unpredictable environments.