Instance-aware adaptive label assignment for 3D object detection

Accurately detecting objects in 3D scenes is crucial for autonomous driving. Although existing point-based methods have made remarkable progress, their performance on small and distant objects remains unsatisfactory. We identify two core issues that contribute to this problem: the imbalance in the number of positive samples and the misalignment between the regression and classification branches. To tackle these challenges, we introduce a single-shot 3D object detector with an instance-aware label assignment strategy and a task-aware balanced loss, named Libra-SSD. Specifically, the instance-aware label assignment strategy employs an adaptive function to collect high-quality samples for instances of varying sizes. Task-aware loss dynamically adjusts the balanced weights for classification and regression tasks based on their learning state. Extensive experiments on the KITTI and nuScenes datasets demonstrate the effectiveness of our method. Our Libra-SSD achieves 63.11% mean average precision (mAP) on the “cyclist" category, while maintaining an outstanding speed of 74 frames per second.