LiDAR Panoptic Segmentation With Learnable Embeddings and Dynamic Decoder

초록

Light detection and ranging (LiDAR) panoptic segmentation is a crucial task for autonomous driving, enabling comprehensive scene understanding by unifying semantic and instance segmentation. However, existing methods face challenges in effectively utilizing available information when constructing positional embeddings (PEs) and adapting to varying scene complexities. This article presents a novel approach to addressing these limitations. We introduce two key innovations: 1) an improved PE technique that uses learnable weighting functions to combine Cartesian, polar, and intensity information, enhancing the ability of the model to distinguish between objects with similar geometries but different material properties and 2) a dynamic decoder depth (DDD) mechanism that adapts to the complexity of each scene, optimizing the tradeoff between accuracy and computational efficiency. Extensive experiments on the SemanticKITTI dataset demonstrate that our approach outclasses state-of-the-art methods by ranking first in six of the ten metrics, achieving superior panoptic quality (PQ) on the test set benchmark while maintaining competitive inference speeds. Our ablation studies confirm the effectiveness of each component and highlight the robustness of our approach across various parameter settings and scene conditions. This study represents an advancement in LiDAR panoptic segmentation, offering improved performance and efficiency for real-world autonomous driving applications.

키워드