Revisiting 3D Object Detection From an Egocentric Perspective Supplementary Material

This document provides supplementary content to the main paper. A, we expand the discussion on comparing our egocentric metric with a recent planner-based 3D object detection metric. B, we provide more details of the StarPoly architecture and training. D shows more visualization results. It measures the KL-divergence of the planner's prediction based on either the ground truth or the detection.