An innovative point cloud compression technology (TSC-PCAC) developed by a joint team from the Chinese Academy of Sciences, Tongji University, and Ningbo University has achieved significant breakthroughs. This technology not only greatly enhances the compression efficiency of point cloud data but also significantly reduces processing time, clearing technical barriers for the development of 3D applications such as AR/VR.
In the context of the rapid development of 3D visual technology, point clouds, as a key data format for virtual and augmented reality, face enormous challenges in transmission and storage. A high-quality point cloud may contain millions of data points, each carrying multi-dimensional information such as position, color, and transparency. The processing efficiency of this massive data directly impacts the adoption speed of 3D applications.
To address this challenge, the research team developed a point cloud attribute compression technology (TSC-PCAC) based on an end-to-end voxel Transformer and sparse convolution. The core of this technology lies in its unique two-stage compression architecture: the first stage focuses on extracting and modeling local features of the point cloud, while the second stage captures global features through a larger receptive field, effectively reducing data redundancy.
The research team also innovatively designed a channel context module based on TSCM, significantly improving data compression efficiency by optimizing the correlation between channels. Experimental data shows that compared to existing mainstream technologies, TSC-PCAC achieved a significant improvement in data compression rates: a 38.53% increase over Sparse-PCAC, a 21.30% increase over NF-PCAC, and an 11.19% increase over G-PCC v23. More impressively, its processing speed also saw a qualitative leap, with encoding and decoding times reduced by 97.68% and 98.78%, respectively.
This groundbreaking achievement not only addresses key pain points in point cloud data processing but also lays an important foundation for the further development of 3D applications such as AR/VR. The research team stated that they will continue to explore deep network technologies with higher compression ratios and work towards a unified processing solution for geometry and attribute encoding in the future.
Paper link: https://arxiv.org/html/2407.04284v1
