In context
By late 2023, 3D reconstruction was a bottleneck for robotics perception, with traditional methods slow and imprecise, and neural radiance fields (NeRF) offering quality but at high computational cost. The emergence of 3D Gaussian splatting (3D GS) promised real-time rendering and editable scenes, making it a pivotal development for industrial automation and robotics.
What was reported
A comprehensive survey from Nankai University analyzed 3D GS technology, which represents scenes with tens of thousands of 3D Gaussian primitives and uses tile-based parallel rasterization for efficient rendering. Compared to NeRF, 3D GS achieves real-time rendering and better scene editability, while maintaining high-quality new view synthesis.
The survey detailed the technical framework, including SfM point cloud initialization, Gaussian ellipsoid parameterization, projection, and adaptive density control to handle under- and over-reconstruction. It also compared 3D GS and NeRF across metrics, noting 3D GS's advantages in training and rendering time.
Applications in visual SLAM were highlighted, with pioneering methods like SplaTAM and GS-SLAM integrating 3D GS for dense mapping. Compact representations were developed to reduce redundancy and improve real-time performance, addressing challenges in large-scale environments.
Why it mattered
3D GS offered a practical path to real-time, high-fidelity 3D reconstruction, enabling robots to perceive and interact with environments more efficiently. Its explicit representation facilitated scene editing, crucial for dynamic environments and human-robot collaboration, and opened new possibilities in autonomous driving and generative AI.
“3D GS technology has achieved real-time rendering without affecting reconstruction quality and visual effects.”
Source: 《机器人》期刊 (robot.sia.cn) · Published 2024-09-15 · “基于3D高斯溅射的3维重建技术综述”