JSCC-Aided INR for High-Frequency Detail Preservation in LiDAR

Light Detection and Ranging (LiDAR) sensors generate accurate 3D representations of real-world environments, which are essential for applications of 3D scene understanding. However, the substantial volume of LiDAR data poses significant challenges for efficient compression and transmission. Implicit neural representation (INR) has gained attention for its compact data representation, but its capacity to accurately represent high-frequency details is insufficient when using small models. In this paper, we propose a novel joint source-channel coding (JSCC) scheme that integrates INR with analog residual transmission for high-quality and efficient point cloud transmission. This scheme is designed to compensate for the limited high-frequency representation of INRs by transmitting the unmodeled details as residuals via pseudo-analog modulation. This integrated approach enables continuous reconstruction quality adaptation to varying wireless channel conditions and effectively mitigates the stair-case effect inherent in conventional digital schemes. Evaluations on the KITTI dataset demonstrate that the proposed scheme outperforms conventional and INR-based compression methods in terms of R-D performance and detection quality at low bitrates.