Reflectance Prediction-based Knowledge Distillation for Robust 3D Object Detection in Compressed Point Clouds

TL;DR

This work tackles robust 3D object detection under low-bitrate compressed LiDAR data for V2X collaboration. It introduces Reflectance Prediction-based Knowledge Distillation (RPKD), which augments non-reflectance compressed point clouds with a receiver-side Reflectance Prediction (RP) module and cross-source distillation (CDTS) to transfer detection knowledge from raw to compressed data. The framework relies on Reflectance Cross-Match (RCM) and Reflectance Inner-Match (RIM) to generate labels and two distillation terms (RKD and DKD) to guide learning, achieving consistent gains on KITTI and DAIR-V2X-V across multiple backbones and code rates. Empirical results show improved detection accuracy and robustness with substantial bandwidth savings, highlighting its practical impact on bandwidth-constrained cooperative perception; code will be released publicly.

Abstract

Regarding intelligent transportation systems, low-bitrate transmission via lossy point cloud compression is vital for facilitating real-time collaborative perception among connected agents, such as vehicles and infrastructures, under restricted bandwidth. In existing compression transmission systems, the sender lossily compresses point coordinates and reflectance to generate a transmission code stream, which faces transmission burdens from reflectance encoding and limited detection robustness due to information loss. To address these issues, this paper proposes a 3D object detection framework with reflectance prediction-based knowledge distillation (RPKD). We compress point coordinates while discarding reflectance during low-bitrate transmission, and feed the decoded non-reflectance compressed point clouds into a student detector. The discarded reflectance is then reconstructed by a geometry-based reflectance prediction (RP) module within the student detector for precise detection. A teacher detector with the same structure as the student detector is designed for performing reflectance knowledge distillation (RKD) and detection knowledge distillation (DKD) from raw to compressed point clouds. Our cross-source distillation training strategy (CDTS) equips the student detector with robustness to low-quality compressed data while preserving the accuracy benefits of raw data through transferred distillation knowledge. Experimental results on the KITTI and DAIR-V2X-V datasets demonstrate that our method can boost detection accuracy for compressed point clouds across multiple code rates. We will release the code publicly at https://github.com/HaoJing-SX/RPKD.

Figures (8)

• Figure 1: Comparison of 3D object detection methods for compressed LiDAR point clouds. (a) With reflectance encoding, detection suffers from the transmission burden of the attribute bitstream. (b) Without reflectance encoding, only geometric coordinates are compressed, improving efficiency but showing poor robustness under a single-source training strategy (STS). (c) Building on raw and non-reflectance compressed data, SparseKD (CDTS) transfers detection knowledge from raw to compressed data detectors, yielding slight performance gains. (d) Our RPKD (CDTS) introduces the RP module and RKD constraint to recover reflectance, significantly boosting detection performance on non-reflectance compressed data.
• Figure 2: Quantitative comparison of different detection methods on multiple KITTI compressed data. STS-R and STS-C denote single-source training strategies based on raw and compressed data, respectively, while CDTS refers to the cross-source distillation training strategy proposed in this paper. The blue dotted line shows the detection performance of the PV-RCNN baseline on raw data.
• Figure 3: Overview of our method. During student detector training, the RCM module assigns reflectance labels from raw point-cloud voxels to compressed points, while the RP module extracts geometric features to predict compressed point reflectance. In contrast, the teacher detector’s RP module predicts raw point reflectance based on labels obtained from the RIM module. The predicted reflectance is then integrated into non-reflectance voxels for detection learning. Building on CDTS, RKD transfers guided reflectance knowledge to the student detector, while DKD utilizes first-stage proposals to convey accurate detection knowledge.
• Figure 4: Two matching methods of the RCM module. All voxels originate from voxelizing raw point clouds. If a compressed point lies within a non-empty voxel, the voxel’s mean reflectance is assigned to it. Otherwise, we search for the surrounding non-empty voxels within a 1-voxel matching range and use the nearest voxel’s mean reflectance as the matching value for this compressed point.
• Figure 5: Principle of CDTS. Dotted lines indicate specialized designs for compressed point-cloud reflectance prediction in RPKD (CDTS), while the red line represents the pipeline for evaluating final performance.