Neural Distribution Prior for LiDAR Out-of-Distribution Detection
Summary: arXiv:2604.09232v1 Announce Type: cross
Abstract
LiDAR-based perception is critical for autonomous driving due to its robustness to poor lighting and visibility conditions. Yet, current models operate under the closed-set assumption and often fail to recognize unexpected out-of-distribution (OOD) objects in the open world. Existing OOD scoring functions exhibit limited performance because they ignore the pronounced class imbalance inherent in LiDAR OOD detection and assume a uniform class distribution.
Introduction
To address this limitation, researchers have proposed the Neural Distribution Prior (NDP), a novel framework aimed at enhancing OOD detection capabilities for LiDAR systems. NDP models the distributional structure of network predictions and adaptively reweights OOD scores based on alignment with a learned distribution prior. This approach seeks to overcome the challenges posed by class imbalances and improve the reliability of autonomous driving systems.
Key Features of the Neural Distribution Prior
- Dynamic Logit Distribution Capture: NDP dynamically captures the logit distribution patterns of training data, allowing for a more accurate representation of the model’s confidence in its predictions.
- Attention-Based Module: An attention-based module corrects class-dependent confidence bias, ensuring that the model’s predictions are more reliable across various classes.
- Perlin Noise-Based OOD Synthesis: The introduction of a Perlin noise-based OOD synthesis strategy generates diverse auxiliary OOD samples from input scans, enabling robust OOD training without the need for external datasets.
Experimental Results
Extensive experiments conducted on the SemanticKITTI and STU benchmarks demonstrate the effectiveness of the NDP framework. Notably, NDP achieves a point-level Average Precision (AP) of 61.31% on the STU test set, which is more than 10 times higher than the previous best result. This significant improvement underscores the potential of NDP in enhancing the performance of LiDAR systems in detecting OOD objects.
Compatibility and Future Work
The NDP framework is designed to be compatible with various existing OOD scoring formulations, providing a versatile and effective solution for open-world LiDAR perception. Future research may focus on further refining the NDP approach and exploring its applicability in real-world scenarios, potentially leading to safer and more reliable autonomous driving systems.
Conclusion
The introduction of the Neural Distribution Prior represents a significant advancement in LiDAR OOD detection, addressing critical challenges in the field of autonomous driving. By modeling distributional structures and correcting biases, NDP enhances the robustness of perception systems, paving the way for improved safety and functionality in real-world applications.