3D Point Cloud Compression


Among various data modalities such as audio, image, text, etc., 3D data in the form of point clouds constitutes a growing portion of the spectrum. Sensors for capturing 3D point cloud data such as light detection and ranging, stereo, structured light, and time-of-flight have become increasingly popular and economical. In comparison to images, high-dimensional information can be described in 3D with immunity to variations in color, illumination, and scale. Besides point clouds, there are numerous ways to depict 3D data including meshes, CAD models, and volumetric representations. However, when compared to the other data representations, point clouds offer the advantage of providing a simpler, denser, and closer-to-reality description. 3D sensors generate enormous amounts of point cloud data at high frame rates. For example, a 3D point cloud with 0.7 million points per frame at 30 frames per second needs a bandwidth of approximately 500 MB/s for video. Working with uncompressed point cloud data can lead to congestion and delays in communication networks. Therefore, efficient compression coding technologies are indispensable for ensuring the compact storage and transmission of 3D point cloud data.


  • We introduced a weighted entropy loss function and inference strategy to compress raw 3D point clouds at different bitrates using a single trained model along with benchmarks for a variety of perception tasks on publicly available datasets


  1. M.A.A Muzaddid and W.J. Beksi, "Variable Rate Compression for Raw 3D Point Clouds," International Conference on Robotics and Automation (ICRA), 2022.
    Paper  •   Preprint  •   Source Code  •   Citation