Point Cloud Rendering failure by "diff_gaussian_rasterization import GaussianRasterizer as Renderer"

[Xiaoyuan-Isaac-Wang]

Hi, I try to use from diff_gaussian_rasterization import GaussianRasterizer as Renderer to render the predicted .ply file, but I consistently encountered the CUDA issue RuntimeError: CUDA error: an illegal memory access was encountered. After I render the ply into image, the im, cam, gs_data will point to some strange address like this:

I think the correct one after rendering should be like this:

Do you have any sense why this would happen? Thanks!

[btsmart]

Hello @luminousking,

Unfortunately I am not sure what could be causing the illegal memory accesses. In my experiments (and in the README) I used this fork of diff_gaussian_rasterization from pixelSplat: https://github.com/dcharatan/diff-gaussian-rasterization-modified, so you might like to check that it is this version you have installed. I conducted by experiments at 256x256 and 512x512 resolution, so you may also wish to verify that your resolution isn't causing any issues.

In my experiments, I think I also never directly render the .ply files that I create, I think I only use those for exporting and visualizing in interactive viewers. If you are using diff_gaussian_rasterization on data from the .ply files, your pipeline may be a little different to mine.

If you are able to find the issue, please let me know!

[EhanDong]

嗨，我尝试使用from diff_gaussian_rasterization import GaussianRasterizer as Renderer来渲染预测.ply文件，但我一直遇到 CUDA 问题RuntimeError: CUDA error: an illegal memory access was encountered。将 ply 渲染成图像后，im、cam、gs_data 将指向一些奇怪的地址，如下所示： 我认为渲染后的正确地址应该是这样的： 你知道为什么会发生这种情况吗？谢谢！

Hello, I tried to migrate this code to scene reconstruction from multiple perspectives, and I found that I also encountered the same problem. Did you solve this problem later?

[brandonsmartdev]

Unfortunately I have not been able to recreate this issue, so I am still not sure what the underlying issue is. If diff_gaussian_rasterization does not work, you may have better success with GSplat, but that will require modifications to the current code. If you are able to make any progress, please let me know.