NRSeg: Noise-Resilient Learning for BEV Semantic Segmentation via Driving World Models
TLDR
NRSeg proposes a noise-resilient learning framework for BEV semantic segmentation using synthetic data from driving world models.
Reasoning
The paper addresses a practical issue (generation noise in synthetic data) and introduces novel components (PGCM, BiDPP, HLSE). However, it is limited to BEV segmentation and does not explicitly validate on real-world benchmarks in the abstract.
Read-first score
Read-first score 59.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 25.
Field roles
Rank sensitivity
Stability: volatile; rank range: 599.
Keyword Scores
Deep Analysis
Innovations
- Perspective-Geometry Consistency Metric (PGCM) to quantitatively evaluate the guidance capability of generated synthetic data
- Bi-Distribution Parallel Prediction (BiDPP) to enhance model robustness through parallel prediction of multinomial and Dirichlet distributions
- Hierarchical Local Semantic Exclusion (HLSE) module to address non-mutual exclusivity in BEV semantic segmentation
Methodology
NRSeg leverages synthetic data from driving world models to improve BEV semantic segmentation. It introduces a Perspective-Geometry Consistency Metric (PGCM) that measures alignment between perspective road masks and BEV label projections to assess data quality. The framework employs a Bi-Distribution Parallel Prediction (BiDPP) mechanism combining multinomial prediction for semantics and Dirichlet-based evidential learning for uncertainty quantification, along with a Hierarchical Local Semantic Exclusion (HLSE) module to handle non-mutual exclusivity in segmentation labels.
Key Results
NRSeg achieves state-of-the-art performance, yielding the highest improvements in mIoU of 13.8% and 11.4% in unsupervised and semi-supervised BEV segmentation tasks, respectively.