FS$^{2}$D: Fully Sparse Few-Shot 3D Object Detection

Corner cases are a focal issue in current autonomous driving systems, with a significant portion attributed to few-shot detection. Due to the sparse distribution of point cloud data and the real-time requirements of autonomous driving, traditional few-shot detection methods face challenges in direct application to the 3D domain, making it more difficult for outdoor scene 3D detectors to handle corner cases. In this study, we employ fully sparse feature matching and aggregation operations, utilizing meta-learning methods to enhance performance on few-shot categories without increasing network inference parameters. Furthermore, our few-shot research is based on the inherent characteristics of publicly available data without introducing additional categories, allowing for fair comparisons with existing methods. Extensive experiments were conducted on the widely used nuScenes dataset to validate the effectiveness of our method. We demonstrate superior performance compared to the baseline method, especially in handling few-shot categories.