3DSSD: Point-based 3D Single Stage Object Detector

January 2021

tl;dr: Single-stage, anchor-free point-based lidar object detector.

Overall impression

This study brings the point-based lidar object detectors (PointRCNN) to the realtime realm. Almost all previous real-time lidar object detectors are voxel based, such as PIXOR, PointPollars and VoxelNet.

Existing point based lidar object detector have largely four stages

The feature propagation and refinement module takes more than half of the time and thus need to be eliminated. The largest contribution of 3DSSD is the use of feature based furthest point sampling to eliminate the need for feature propagation. –> This reminds me of feature-metric loss ECCV 2020 in monocular depth estimation, and BEV feat sticthing.

The paper introduces quite a few acronyms which makes the paper a bit hard to read.

Key ideas

Technical details