NRSeg: Noise-Resilient Learning for BEV Semantic Segmentation via Driving World Models

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Birds' Eye View (BEV) semantic segmentation is crucial for autonomous driving perception.
• Current unsupervised and semi-supervised methods struggle with homogeneous labeled data distribution.
• Synthetic data offers potential but is hampered by generation noise.

Purpose of the Study:

• To develop a noise-resilient learning framework (NRSeg) for robust BEV semantic segmentation.
• To leverage synthetic data from world models to enhance labeled data diversity.
• To improve the performance of autonomous driving perception systems.

Main Methods:

• Proposed NRSeg framework with Perspective-Geometry Consistency Metric (PGCM) for evaluating synthetic data.
• Implemented Bi-Distribution Parallel Prediction (BiDPP) for model robustness and uncertainty quantification.
• Introduced Hierarchical Local Semantic Exclusion (HLSE) module for non-mutual exclusivity in segmentation.

Main Results:

• NRSeg achieved state-of-the-art performance on BEV semantic segmentation tasks.
• Significant improvements in mean Intersection over Union (mIoU): 13.8% (unsupervised) and 11.4% (semi-supervised).
• Demonstrated effectiveness across multiple world models and on the nuScenes dataset.

Conclusions:

• NRSeg effectively harnesses synthetic data for robust BEV semantic segmentation.
• The proposed framework overcomes limitations of generation noise and data homogeneity.
• NRSeg significantly advances the capabilities of autonomous driving perception systems.