Can We Remove the Ground? Obstacle-aware Point Cloud Compression for Remote Object Detection

TL;DR

The paper addresses the challenge of low-latency, bitrate-efficient point cloud transmission for remote object detection by selectively removing ground points. It proposes a lightweight pillar-based Ground Removal (PGR) method comprising Pillar removal and Pillar restoration, governed by physically meaningful thresholds $δ_{minmax}$, $δ_{env}$, environmental radius $er$, and adaptive $δ_{res}$, with geometry encoded via G-PCC. Empirical results on KITTI and Waymo show that removing about 20-30% of ground points can maintain detection performance (mAP) while reducing bitrate, achieving up to 86 FPS processing. This obstacle-aware compression enhances streaming efficiency for cooperative perception without heavy semantic models and demonstrates robustness to parameter tuning, with avenues for future automatic parameter optimization and improved handling of uneven terrain.

Abstract

Efficient point cloud (PC) compression is crucial for streaming applications, such as augmented reality and cooperative perception. Classic PC compression techniques encode all the points in a frame. Tailoring compression towards perception tasks at the receiver side, we ask the question, "Can we remove the ground points during transmission without sacrificing the detection performance?" Our study reveals a strong dependency on the ground from state-of-the-art (SOTA) 3D object detection models, especially on those points below and around the object. In this work, we propose a lightweight obstacle-aware Pillar-based Ground Removal (PGR) algorithm. PGR filters out ground points that do not provide context to object recognition, significantly improving compression ratio without sacrificing the receiver side perception performance. Not using heavy object detection or semantic segmentation models, PGR is light-weight, highly parallelizable, and effective. Our evaluations on KITTI and Waymo Open Dataset show that SOTA detection models work equally well with PGR removing 20-30% of the points, with a speeding of 86 FPS.

Figures (7)

• Figure 1: Object ground truth (green) and detection results (blue) using PVT-SSD with an input of original (left) PC, PC with semantic ground removal (middle, PolarNet), and PC with PGR (right). Using PC with semantic ground removal, detection bounding boxes are mismatched with ground truth (red circle), while PGR does not affect detection performance.
• Figure 2: Feasibility study on ground point removal for Car detection in KITTI Dataset. The evaluation ranges from uncompressed 'raw' PCs to 'compressed-only' without ground point removal, and then through a series of Extension Factors (EFs) combined with compression.
• Figure 3: Illustration of PGR with reduced number of pillars.
• Figure 4: bpp vs mAP on KITTI dataset using (a) SECOND and (b) PVT-SSD. Dotted lines represent results on uncompressed data. The up arrow $\uparrow$ after mAP means that higher is better and the down arrow $\downarrow$ after bpp means that lower is better.
• Figure 5: bpp vs mAP on Waymo Open Dataset using PVT-SSD. Dotted lines represent results on uncompressed data.