Jiazhong Cen
Jiazhong Cen
你好，nvcc -V输出的版本是cu116吗，这大概率是cuda extension的编译中出现了问题。
我移除了之前在sam3d.py中错误的对mobile_sam库的导入，现在应该可以了
不出现GUI界面应该是由于已存在分割好的文件coarse_segmentation_gui.tar，可以尝试对这个文件重命名，来进行下一次分割。 理论上是可以从这个文件中到处grid的，但需要手动实现一下采样。
有一定的相关性，这里面定义了tensorf grid和普通的3D grids，以及相关的操作。
mask就是用这里的[dense grid](https://github.com/Jumpat/SegmentAnythingin3D/blob/b25698cbc7ca8e71b73749685b36f2914d6cbee7/lib/grid.py#L39)保存的，输入是三维空间中的坐标，输出是对应的三维位置的属性。
Hi. Have you run 'pip install -e .' ?
Hi, this is an unfixed bug in SA3D-GS code. We will fix it within this month. You can also try to remove the involvement of FeatureGaussianModel by yourself.
We use 22.04. You need to align the cuda version with the PyTorch version. You can try to update the PyTorch. 1.12.1 is not necessary.
Hi, have you checked 'nvcc -V' ? The 'CUDA_HOME' and 'LD_LIBRARY_PATH' ? I think maybe torch cannot find the cuda. Does 'torch.cuda.is_available()' ?
Hi, you may check [this](https://discuss.pytorch.org/t/unable-to-find-a-valid-cudnn-algorithm-to-run-convolution/78724/2) for some help. I have never met such problem in sa3d.