SECOND-DX: Single-model Multi-class Extension for Sparse 3D Object Detection

Abstract

3D object detection is becoming increasingly significant for emerging autonomous vehicles. Safety decision making and motion planning depend highly on the result of 3D object detection. Recent 3D detection models are optimized for cars, cyclists and pedestrians with multiple models. This is not desirable because multiple models require significant resources, which are also used for other algorithms, such as localization or object tracking. We present SECOND-DX for providing multi-class support for 3D object detection with only a single model and it enables the detection of all three classes of 3D objects scanned using LiDAR sensors in real time. We conducted experiments involving the KITTI 3D object dataset to show that SECOND-DX is more accuracy overall evaluation metrics without compromising execution speed when compared with algorithms extended to support multi-class detection with a single model. Additionally, SECOND-DX can detect pedestrian classes comparable with that of current models that are optimized to support only cyclists and pedestrians.