Yifan Lu

Hi! I also run the no-collaboration experiment again, and the result turns out to be similar as yours. And In my [checkpoints](https://huggingface.co/yifanlu/HEAL/tree/main) (where I also upload the metric and visualization...

And to be honest, if you infer with `--fusion_method late`, you can get a even higher AP score in DAIR-V2X dataset though there are many false positives.

要注意一下两个版本的Checkpoint是不通用的

其实在写这个仓库代码的时候，我也把opencood对spconv 2.x的支持照抄了过来了。但我本地一直是1.2.1的版本，所以一直没有对2.x的功能进行测试，其他issue里面有人反映一些模型会报错，比如SECOND？如果您在使用2.x的过程中发现里面code有bug，非常欢迎提交PR，你也会成为一个contributor！

Could you try `conda install nvidia/label/cuda-11.8.0::cuda-cupti` ?

It seems that you install cuda via conda, in which case some extra CUDA library would not be installed automatically. The `/usr/bin/ld` can not find the cupti. Installing them manually...

In a system level CUDA (not in the conda environment), they can be found in the `extras` folder

Maybe this helps. https://discuss.pytorch.org/t/build-errors-undefined-ref-to-cuptixxx-solved/183318 Yeah a sure-fire way is to install a local CUDA. it's not a hassle and doesn't require sudo anyway.

You may try to specify those two libraries in the [CMakeLists.txt](https://github.com/yifanlu0227/ChatSim/blob/main/chatsim/background/mcnerf/CMakeLists.txt), so that the cmake can find them explicitly.

sad to hear that, 3dgs is really a good alternative. The installation is much easier