Incorporating Contextual Priors in Autonomous Vehicles
P.B. Chougule (TU Delft - Mechanical Engineering)
Holger Caesar – Mentor (TU Delft - Mechanical Engineering)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Manual annotation of 3D bounding boxes in LiDAR point clouds is a major cost in constructing autonomous driving datasets, motivating offline auto-labelling pipelines. We investigate whether a graph neural network operating on object detector proposals can improve label quality. We construct a radius graph over detections in each nuScenes keyframe and apply a shared GATv2 backbone with classification and regression heads, including a direction classifier. Our joint model improves mAP from 50.25 to 52.48 ± 0.13, NDS from 60.16 to 61.29 ± 0.08, and mAOE from 0.3956 to 0.3783 ± 0.0022 (n = 3) on CenterPoint proposals evaluated on the nuScenes validation split. This improvement arises from rescoring and discrete 180° heading correction.