Neural Distribution Prior Framework Improves LiDAR Out-of-Distribution Detection for Autonomous Driving
A recent study introduces the Neural Distribution Prior (NDP) framework aimed at overcoming the shortcomings of LiDAR-based perception systems utilized in autonomous vehicles. Existing models frequently struggle to identify unexpected out-of-distribution (OOD) objects due to their reliance on closed-set and uniform class distribution assumptions. The NDP framework effectively models the distributional characteristics of network predictions, adaptively reweighting OOD scores in accordance with a learned distribution prior. This method captures logit distribution patterns from training data and mitigates class-dependent confidence biases through an attention-based mechanism. Additionally, the research tackles the significant class imbalance present in LiDAR OOD detection, which current scoring functions overlook. It also presents a Perlin noise-based OOD synthesis technique for generating varied training samples. LiDAR perception is essential for autonomous driving, particularly in challenging lighting and visibility scenarios. This research was published as arXiv:2604.09232v2, categorized under replace-cross.
Key facts
- The Neural Distribution Prior (NDP) framework addresses LiDAR out-of-distribution detection limitations
- Current autonomous driving models operate under closed-set assumptions and fail with unexpected objects
- Existing OOD scoring functions ignore pronounced class imbalance in LiDAR detection
- NDP models distributional structure of network predictions and adaptively reweights OOD scores
- The framework dynamically captures logit distribution patterns from training data
- An attention-based module corrects class-dependent confidence bias
- A Perlin noise-based OOD synthesis strategy generates diverse training examples
- LiDAR perception is critical for autonomous driving due to robustness in poor conditions
Entities
—