麒砺创新技术（深圳）有限公司用户3321
3.12 Press Release: Restoring Reality: 3D Gaussian to Mesh 2.0 (Media Ver.)
输入“/”快速插入内容
3.12 Press Release: Restoring Reality: 3D Gaussian to Mesh 2.0 (Media Ver.)
3.12 Press Release: Restoring Reality: 3D Gaussian to Mesh 2.0 (Media Ver.)
用户4158用户4158
用户3321用户33212024年11月27日修改
Exactly a year ago, KIRI Engine released 3D Gaussian Splatting across all supported platforms, two months later its built-in editing tools, and another two months later 3DGS to Mesh. This brand new 3D scanning method has received lots of utilization and love as optimizations and tools dedicated to it are continously introduced, making this seemingly 'too-good-to-be-true' 3D visualization technique more accessible and practical.
Today, a year later, KIRI Engine is excited to announce two revolutionary updates in its 3.12 release: 3D Gaussian Splatting to Mesh 2.0 and KIRI Engine 3DGS Blender Addon V2.0. Along with other quality of life changes and improvements to the overall versatility of the app, this is an update you don't want to miss.
3DGS (3D Gaussian Splatting) to Mesh 2.0
In January 2024, KIRI released 3DGS to Mesh 1.0. This was made possible in collaboration with CJ Ye and his team, and it was the first attempt to make 3DGS more applicable in widely-used applications as the native PLY point cloud file was hardly accepted and visualized correctly in any rendering engine. This method worked by running a separate algorithm on top of the existing splat, reconstructing surface meshes from the depth map analyzed from the generated gaussian splat result. While this method worked to an extent, the light and depth data was often inaccurate and resulted largely inconsistent results, with featureful objects being converted into mesh files with much higher fidelity than their featureless counterparts.
Figure 1: 3DGS to Mesh 1.0 Generated Suitcase
Building on top of the solid foundation provided by the lastest gaussian splatting mechanics, the newly updated gaussian splat-to-mesh conversion introduces sophisticated normal prediction and reflection removal techniques. Developed once again by the KIRI Engine team and CJ Ye, the new algorithm draws inspiration from professional 3D scanners, specifically the super-dense point cloud generation methodology. With this change, the algorithm is now able to handle scenarios that used to be impossible for gaussian splat-to-mesh reconstruction, which were scanning reflective and transparent surfaces as they would disrupt the reconstruction process by providing inaccurate depth data.
This update has been uploaded on GitHub, along with model viewers that display the differences between 1.0 and 2.0. The model viewers are rigged to be in synch in terms of the camera position, making the models' differences stand out clearly.
Figure 2: 3DGS-to-Mesh 2.0 versus 1.0 on GitHub