If you have great ideas,
Let's talk!

blog

看paper (3d gaussian splatting for real-time radiance field rendering)

3d相关export

https://github.com/graphdeco-inria/gaussian-splatting

看这两篇 都是很新的技术

https://arxiv.org/abs/2410.20030

https://arxiv.org/abs/2308.04079

https://research.nvidia.com/labs/toronto-ai/xcube/

这里顺便对3D重建方式做一个小小的总结

SFM(Structure from Motion), MVS(Multi View Stereo)

Key point detection, matching(sparse point cloud) →

根据特征点推出一般点的位置 → dense point cloud

Neural Field(NERF, Neural SDF(这个也算是neural)))

只要是用NN拟合

Nerf 得到的是nerual representation 图片性质

通过extraction 才能得到模型(点云 mesh voxel…)

训练时间长

Gaussian Splatting

把点变成gaussian

几何形状不好 存储高