Point Cloud Feature Coding for Object Detection over an Error-Prone Cloud-Edge Collaborative System

A research team from institutions including City University of Hong Kong and University of Leicester has published a breakthrough paper on arXiv titled 'Point Cloud Feature Coding for Object Detection over an Error-Prone Cloud-Edge Collaborative System.' The work addresses a critical bottleneck in distributed AI systems: efficiently transmitting 3D perception data between resource-constrained edge devices (like vehicle sensors) and powerful cloud servers. Their framework combines task-driven compression with adaptive error correction to enable reliable object detection even over noisy wireless channels where traditional methods fail.

The technical approach uses a lightweight feature compaction module that identifies and removes task-irrelevant regions from multi-scale point cloud representations, then applies channel-wise dimensionality reduction to preserve only essential information. For transmission, it employs SNR-adaptive channel coding for attribute data and Low-Density Parity-Check (LDPC) encoding for geometric information. On the receiving cloud side, a diffusion-based feature upsampling module reconstructs the complete multi-scale features. When tested on the KITTI autonomous driving dataset, the system achieved remarkable 3D average precision scores of 93.17%, 86.96%, and 77.25% for easy, moderate, and hard objects respectively—all while compressing features 172-fold and operating over challenging 0 dB SNR wireless conditions. The team will release their source code on GitHub, potentially accelerating deployment of efficient edge-cloud perception systems.