Point Cloud Compression With Adaptive Downsampling and Upsampling

The patent segments the point cloud processing into distinct downsampling and upsampling stages. The downsampling stage extracts low-frequency information by subsampling points, while the upsampling stage reconstructs high-frequency details. This segmentation allows each stage to be optimized independently, improving overall coding performance while managing computational complexity through structured processing steps.

The patent applies preliminary downsampling before compression to reduce the number of points and lower the bit rate. By performing this action in advance, the system prepares the point cloud for more efficient encoding while preserving essential low-frequency information that can be recovered during upsampling, thus improving coding performance without requiring all original points to be processed at full resolution.

The patent merges the training of downsampling and upsampling networks into a single unified training process. Both networks are trained simultaneously using combined loss functions that enforce consistency between the downsampled and original point clouds. This merging eliminates the reconstruction errors that occur when processes are trained separately, as the joint training ensures coherent transformation throughout the entire pipeline while maintaining defined processing steps.

The patent implements feedback mechanisms through loss functions that compare the upsampled reconstruction with the original point cloud. The combined loss functions provide continuous feedback during training to adjust both downsampling and upsampling operations, ensuring that the reconstruction quality is optimized. This feedback loop resolves reconstruction errors by iteratively improving the consistency between input and output representations.