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

Cloud-edge collaboration enhances machine perception by combining the strengths of edge and cloud computing. Edge devices capture raw data (e.g., 3D point clouds) and extract salient features, which are sent to the cloud for deeper analysis and data fusion. However, efficiently and reliably transmitting features between cloud and edge devices remains a challenging problem. We focus on point cloud-based object detection and propose a task-driven point cloud compression and reliable transmission framework based on source and channel coding. To meet the low-latency and low-power requirements of edge devices, we design a lightweight yet effective feature compaction module that compresses the deepest feature among multi-scale representations by removing task-irrelevant regions and applying channel-wise dimensionality reduction to task-relevant areas. Then, a signal-to-noise ratio (SNR)-adaptive channel encoder dynamically encodes the attribute information of the compacted features, while a Low-Density Parity-Check (LDPC) encoder ensures reliable transmission of geometric information. At the cloud side, an SNR-adaptive channel decoder guides the decoding of attribute information, and the LDPC decoder corrects geometry errors. Finally, a feature decompaction module restores the channel-wise dimensionality, and a diffusion-based feature upsampling module reconstructs shallow-layer features, enabling multi-scale feature reconstruction. On the KITTI dataset, our method achieved a 172-fold reduction in feature size with 3D average precision scores of 93.17%, 86.96%, and 77.25% for easy, moderate, and hard objects, respectively, over a 0 dB SNR wireless channel. Our source code will be released on GitHub at: https://github.com/yuanhui0325/T-PCFC.

💡 Research Summary

The paper addresses the critical challenge of transmitting high‑dimensional point‑cloud features from edge devices to the cloud for 3‑D object detection in noisy wireless environments. Existing point‑cloud compression standards (G‑PCC, V‑PCC) are designed for human visual quality and are ill‑suited for machine‑vision tasks, while most prior machine‑oriented compression methods either require full reconstruction of raw point clouds or ignore channel‑induced errors. To bridge this gap, the authors propose a joint source‑channel coding framework called T‑PCFC (Task‑driven Point Cloud Feature Coding).

The method first extracts the deepest layer of a multi‑scale feature pyramid produced by a voxel‑based detector (VirConv‑L). A lightweight Feature Compaction module then separates geometry and attribute information. The geometry compaction block, built on a U‑Net architecture with a novel point‑wise loss, removes task‑irrelevant points while preserving regions critical for detection. In parallel, an attribute compaction block employs channel‑wise attention and sparse convolutions to prune redundant feature channels, achieving a drastic reduction in feature size.

Transmission is performed with unequal protection. Geometry data, which is highly sensitive to small errors, is encoded with a traditional Low‑Density Parity‑Check (LDPC) code to guarantee strong error correction. Attribute data, on the other hand, is sent through a deep learning‑based, SNR‑adaptive Joint Source‑Channel Coding (JSCC) encoder that dynamically adjusts the compression rate and scaling according to the instantaneous signal‑to‑noise ratio. At the cloud side, an LDPC decoder restores geometry, while an SNR‑adaptive JSCC decoder reconstructs the attributes.

After decoding, a Feature Decompaction step restores the original channel dimensionality, and a diffusion‑based Feature Upsampling module, guided by a prompt generation block, synthesizes shallow‑layer features from the compact representation. This diffusion model effectively performs multi‑scale reconstruction without transmitting the full‑resolution feature maps, thereby saving bandwidth while preserving detection performance.

Experiments on the KITTI benchmark under a harsh 0 dB SNR wireless channel demonstrate a 172‑fold reduction in transmitted feature size. Despite this extreme compression, the system achieves 3‑D average precision (AP) of 93.17 % (Easy), 86.96 % (Moderate), and 77.25 % (Hard), outperforming direct transmission of compressed features, especially in low‑SNR regimes. The proposed pipeline also respects edge‑device constraints, as the compaction modules are lightweight and suitable for real‑time operation.

Key contributions include: (1) a novel geometry‑aware and channel‑aware feature compaction strategy that simultaneously reduces data volume and improves feature quality; (2) an unequal protection scheme that combines LDPC for geometry and adaptive JSCC for attributes, delivering robustness against wireless noise; (3) a diffusion‑based upsampling mechanism that reconstructs multi‑scale features from compact codes, eliminating the need for full‑scale transmission.

Overall, the work presents the first integrated source‑channel coding solution tailored to point‑cloud feature transmission for machine perception, offering a practical, efficient, and robust foundation for cloud‑edge collaborative applications such as autonomous driving, robotics, and smart‑city surveillance.