This paper introduces a Joint Geometry-Semantic Compression (JGSC) framework to efficiently compress 3D point cloud data while maintaining geometric fidelity and semantic information. The method uses a unified variational autoencoder with cross-modal attention to encode geometry and semantics into a shared latent space, and a task-aware rate-distortion optimization objective to balance reconstruction quality and downstream segmentation performance.
Key findings
Proposes a novel JGSC framework that jointly compresses geometry and semantics in 3D point clouds.
Introduces a unified variational autoencoder architecture with cross-modal attention mechanisms.
Presents a task-aware rate-distortion optimization objective for balancing reconstruction quality and segmentation performance.
Expected to achieve 40-60% bitrate reduction over conventional approaches while maintaining segmentation mIoU within 2% of uncompressed inputs.
Limitations & open questions
The proposed method's performance in real-world applications and scalability to larger datasets is yet to be evaluated.
The framework's sensitivity to hyperparameter tuning and potential impact on compression efficiency needs further investigation.