LetianHuang / op43dgs

[ECCV'24] On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy
https://letianhuang.github.io/op43dgs
Other
97 stars 6 forks source link

Use pinhole to train? #5

Open haofengsiji opened 3 weeks ago

haofengsiji commented 3 weeks ago

Is the process using a pinhole to train, and then the image can be rendered as a pinhole, fisheye ..., like this?

LetianHuang commented 3 weeks ago

Yes, after training with the pinhole model, the scene can be adapted to different camera models by replacing the submodules of diff-gaussian-rasterization-(pinhole,panorama, ...), enabling the rendering of pinhole, fisheye, panorama, and other types of images. If the backward propagation code backward.cu for the corresponding camera model has been implemented, it is also possible to use pinhole, panorama, fisheye, and other models' images as inputs to train the scene, since the optimal projection proposed in our paper is differentiable.