search

ShengYu724 / MSGField

1 stars 0 forks source link

Question about the paper #2

Open zcc00210 opened 4 days ago

zcc00210 commented 4 days ago

Thanks for your wonderful work! I have a question about the content in this article. You mentioned in Sec.IV E about the effects of object-centered feature distillation that Langsplat performs poorly in 3d segmentation while 2d results are excellent due to the fact that features lose meaning after α-blending. Based on this, what I don't quite understand is, doesn't α-blending happen in the process of rendering 3D Gaussian distributions into 2d images, and if blending features would make it meaningless, then wouldn't it make sense that the 2D result is bad (because α-blending occurs) and the 3d result is good? Looking forward to your reply！

ShengYu724 commented 3 days ago

Thank you for your attention! Alpha-blending can be simplified as: ax + (1-a)y = z. The training process ensures that the rendered results (z) match the ground truth (z′), which represents the compressed 2D feature. While the rendered result (z) is accurate, the features on the 3D Gaussian splats (x or y) cannot be guaranteed to be as meaningful as z. I hope this answers your question.