search

TY424 / AbsGS

Other
137 stars 4 forks source link

Train Problem #9

Open suvzzc opened 2 months ago

suvzzc commented 2 months ago

图片 Thank you for your excellent work. I tried to move your work to MonoGS as its renderer, but I had certain problems. I would be very grateful if you could help me!

TY424 commented 1 month ago

Hello, I haven't used MonoGS before. The dL_dmean2D in 3DGS was of shape N×3 (only using the first two dimensions). We changed it to N×4, keeping the first two dimensions the same and allocating the last two dimensions with Homodirectional Gradient. You can refer to backward.cu for modifications, and be sure to update the corresponding function parameters.

CosyXu commented 1 month ago

I have met the same error when transferring this great work to another 3DGS method. I think it is caused by the 4-shape of the dL_dmean2D, and how to solve it?

TY424 commented 1 month ago

I have met the same error when transferring this great work to another 3DGS method. I think it is caused by the 4-shape of the dL_dmean2D, and how to solve it?

@suvzzc

The corresponding variables need to be changed to 4D. For example:

2024-10-11_10-30

suvzzc commented 1 month ago

I have met the same error when transferring this great work to another 3DGS method. I think it is caused by the 4-shape of the dL_dmean2D, and how to solve it?

@suvzzc

The corresponding variables need to be changed to 4D. For example:

2024-10-11_10-30

Thanks, I've got it running successfully.I would also like to ask something about the reduction of Gaussian floating objects. In running MonoGS, I noticed that there is a large number of floating 3D Gaussian ellipsoids behind it and it doesn't seem to be significantly reduced after I added your code for reducing the floating objects, can't the floating objects here be reduced by setting the transparency? I would be very grateful if you could help me! 图片