Closed KzZheng closed 5 months ago
I also encountered this issue. Have you solved it now?
I suppose it is the problem of modified cuda_rasterizer/backward.cu:495. When include_feature=False, language_feature is (1,0), and this line tries to access this variable as if its the language_feature (N_gaussian, dim_feature)
I just use the original 3DGS to get the original point cloud.
@KzZheng @zqh0253 @xuqinwang. Have a look at https://github.com/minghanqin/langsplat-rasterization/pull/1 to see if it fixes your issue.
Another hacky workaround I found while looking for the bug is to override https://github.com/minghanqin/LangSplat/blob/850e8b94bddacb0ad1d8173cc43793a0f0338112/gaussian_renderer/__init__.py#L50 to include_feature=True,
, and https://github.com/minghanqin/LangSplat/blob/850e8b94bddacb0ad1d8173cc43793a0f0338112/gaussian_renderer/__init__.py#L91 to language_feature_precomp = colors_precomp * 0.0
(with --convert_cov3D_python
flag).
Thanks a lot! @nhat-vo I have merged your modifications into the original code.
@KzZheng @zqh0253 @xuqinwang. Have a look at minghanqin/langsplat-rasterization#1 to see if it fixes your issue.
Another hacky workaround I found while looking for the bug is to override
to
include_feature=True,
, and https://github.com/minghanqin/LangSplat/blob/850e8b94bddacb0ad1d8173cc43793a0f0338112/gaussian_renderer/__init__.py#L91to
language_feature_precomp = colors_precomp * 0.0
(with--convert_cov3D_python
flag).
Thnak you so much !
Hi,
I have tried to set include_feature=False for the original 3DGS training, but I encounter an error after the model runs several iterations:
I can successfully train with the original 3DGS code.