First time streaming in the U.S. (Talking about Gaussian Splatting cuda code) [English CC]

Рет қаралды 12,955

10 ай бұрын

note written during the stream: app.conceptboard.com/board/mi...
So I came to visit luma office, and want to share my feeling during the stay, also talk about the work we're doing recently

Пікірлер: 44

@user-cx4cq8nx1f 7 ай бұрын

感谢分享终于找到一个能听懂的视频了！

@user-ir3ql5yk7y 9 ай бұрын

非常棒的讲解，从nerf就开始关注你啦~~

@fangliu125 9 ай бұрын

我真的太爱葵大了😍向葵大学习！

@jiashuaizmo9444 8 ай бұрын

Thanks a lot for these sharing about 3DGS. I am working on this copy of code currently.

@silbarlvel 3 ай бұрын

感谢葵大，讲的太好了，反复刷了好几遍每次都有新收获，希望以后能看到更多的教程

@duckbreeder 9 ай бұрын

謝謝！

@marsen1106 6 ай бұрын

thanks a lot for the explaining

@tonyzhang2501 6 ай бұрын

非常感谢

@tianchensong4081 4 ай бұрын

谢谢，讲解的非常清楚

@tianchensong4081 4 ай бұрын

不知道是否可以再讲讲如果有球谐函数的反向传播怎么计算，直播里只讲了render这一步的反向传播，preprocess的感觉更复杂不过讲的部分非常清晰非常感谢

@panicnever6683 8 ай бұрын

太感谢了！！！

@zichenwang8068 9 ай бұрын

超级感谢葵大！当初就是看您的 NeRF 讲解入门，如今想学习 3dGS，还是要看您的视频，多谢分享~

@asdfasdfsd 7 ай бұрын

我最近也刚开始学习NeRF，零基础，要怎么入门呢

@bnext594 3 ай бұрын

感谢分享

@gorilla543210 4 ай бұрын

十分感謝！有個問題想問，目前3DGS是用undistort好的image去做，但如果是用distort的image去做3DGS，不知道可不可以把3D->2D的過程加上distortion校正後就可以正常渲染？

@pooip2169 6 ай бұрын

請問以下葵大與各位朋友幾個問題我在嘗試建立重建人的時候, 往往臉的效果都特別差這有啥拍攝手法或系統中參數調整可以優化嘛？另外一問有沒有這方面的論壇能夠跟大家交流學習的地方.

@user-kn3bo5sj2n 8 ай бұрын

If you have time, I would also like to hear from you about Python code!!😭

@cxzhou 4 ай бұрын

什么时候讲instant-ngp的cuda代码，葵葵

@user-vr7xw1qc1p 3 ай бұрын

请问有没有办法可以筛选出render时用到的gaussian，把没有用到的gaussian都给删掉呢

@RDY_97 21 күн бұрын

你是如何在 mac 上用鼠标滚动得如此平滑的，是什么特殊的鼠标吗

@user-uu7jj9sj7b 8 ай бұрын

那这个和slam结合如何

@er-wl9sy 9 ай бұрын

Turning on English subtitles will help

@kwea123 9 ай бұрын

@user-zy6pt8li7x 8 ай бұрын

葵大，你好。如果我想要run或者debug这个代码，该如何操作呢？在您的cuda教学视频里，似乎也没有提到如何直接run写好的C++程序，而是直接使用pip安装。

@xuchen-kw4dm 6 ай бұрын

你好，我也有这个问题。请问有解决的办法了吗

@RDY_97 21 күн бұрын

python 调用 cuda/c++ 实现，光栅化实现是单独的git仓库。在主仓库 submodules 目录下面。

@user-vz4zz5od2g 7 ай бұрын

感谢葵大分享！真的学到很多，然后有一个我在实际操作中遇到的问题，不知道葵大能否指点解惑一下，就是我最近也在训练这个高斯溅射的模型，但是最后训练出来的效果总感觉没有官方的效果那么好，我用来训练的照片分辨率也都是5K的，数量大概200-300张，无论是分辨率还是数量应该都远远大于官方训练的素材，然后我发现官方给出的他们最终的训练结果只是用了30几张图片，但是最后训练出来的文件大小是1个多G的大小，但是我无论怎么训练最后生成的文件大小始终不超过1个G，我觉得这个可能是我和官方质量上有差距的原因，但是不知道到底是什么差在了哪个步骤导致的，您在研究时有遇到过这种情况吗？该如何才能训练出来效果最好的高斯模型呢？

@sophist52 7 ай бұрын

還是你Gaussian的數量丟的不夠多？

@user-vz4zz5od2g 7 ай бұрын

您指的Gaussian是什么的数量？点云的数量吗？我是用的metashape获取的位姿，然后丢给高斯进行计算的@@sophist52

@bookerml8502 7 ай бұрын

@@user-vz4zz5od2g 初始化的点云数量不够，我做室内场景发现有个问题是如果你初始化点太少，或者不在关键位置上，效果不好。我感觉你把图片丢到colmap做dense reconstruction，然后把dense points丢进去初始化，估计效果能不错

@user-vz4zz5od2g 7 ай бұрын

好的！我尝试一下！感谢分享！@@bookerml8502

@JaswinKasi 6 ай бұрын

Hi, given a pixel, is there a function to know what all Gaussian splat is contributing to it ?

@kwea123 5 ай бұрын

the original code already does that, but it stores in some buffer that is very hard to extract from pytorch side. If you want to retrieve that info, you basically need to create a pytorch tensor, and pass the pointer to that tensor to the cuda code, then you can retrieve from python. It is rather complicated.

@JaswinKasi 5 ай бұрын

@@kwea123can you tell me which part of the code does it. I dont know CUDA much

@kwea123 5 ай бұрын

@@JaswinKasi github.com/graphdeco-inria/gaussian-splatting/issues/403 I pointed out in this issue, but it is rather complicated without code explanation.

@JaswinKasi 5 ай бұрын

@@kwea123 thanks!