An Empirical Analysis of Range for 3D Object Detection
This addresses a critical gap in autonomous navigation by enabling detection of far-field objects for safer planning, though it is incremental as it builds on existing methods with dataset-specific tuning.
The paper tackled the problem of far-field 3D object detection for autonomous vehicles, which is overlooked in benchmarks, and found that using large voxels for sparse far-field LiDAR data and ensembling range experts improved efficiency by 33% and accuracy by 3.2% CDS.
LiDAR-based 3D detection plays a vital role in autonomous navigation. Surprisingly, although autonomous vehicles (AVs) must detect both near-field objects (for collision avoidance) and far-field objects (for longer-term planning), contemporary benchmarks focus only on near-field 3D detection. However, AVs must detect far-field objects for safe navigation. In this paper, we present an empirical analysis of far-field 3D detection using the long-range detection dataset Argoverse 2.0 to better understand the problem, and share the following insight: near-field LiDAR measurements are dense and optimally encoded by small voxels, while far-field measurements are sparse and are better encoded with large voxels. We exploit this observation to build a collection of range experts tuned for near-vs-far field detection, and propose simple techniques to efficiently ensemble models for long-range detection that improve efficiency by 33% and boost accuracy by 3.2% CDS.