Transformer and Upsampling-Based Point Cloud Compression
Junteng Zhang, Gexin Liu, Dandan Ding, Zhan Ma
Abstract
Learning-based point cloud compression has exhibited superior coding performance over the traditional methods such as MEPG G-PCC. Considering that conventional point cloud representation formats (e.g., octree or voxel) will introduce additional errors and affect the reconstruction quality, we directly use the point-based representation and develop a framework that leverages transformer and upsampling techniques for point cloud compression. To extract latent features that well characterize an input point cloud, we build an end-to-end learning framework: at the encoder side, we leverage cascading transformers to extract and enhance useful features for entropy coding; At the decoder side, in addition to the transformers, an upsampling module utilizing both coordinates and features is devised to reconstruct the point cloud progressively. Experimental results demonstrate that the proposed method achieves the best coding performance against state-of-the-art point-based methods, e.g., >1 dB D1 and D2 PSNR at bitrate 0.10 bpp and more visually pleasing reconstructions. Extensive ablation studies also confirm the effectiveness of transformer and upsampling modules.