RGB_FILLED failed but depth successed

[rainprob]

[rainprob]

I tried RGB_PREFILLED, NORMAL, and SEMENTICS.
RGB_PREFILLED can display the correct content but it still has some wrong pixels, the results of NORMAL and SEMENTICS are both black. _

[rainprob]

I tried all solutions in issues, such as checking the version of torchvision and torch, kill depth_render process, and others. But these all don't work.

[Berk035]

Hello @rainprob ,

I recommend to reinstall the GibsonEnv clearly. You can try to install docker version instead of building source. Have a nice day.

[fxia22]

@rainprob I observe similar issues when the gibson/core/render/render_cuda_f.so is not correctly compiled with nvcc. Can you check your nvcc --version is the same as nvidia-smi version?

[rainprob]

@fxia22
You are right. My nvcc --version is 11.0, however nvidia-smi version is 11.2. So, how can I solve that? Make them both versions same？

[rainprob]

@Berk035
Thanks for your reply.

[rainprob]

@fxia22
You are right. My nvcc --version is 11.0, however nvidia-smi version is 11.2. So, how can I solve that? Make them both versions same？

In gibson docker container, nvcc --version is 9.0 and nvdia-smi version is 11.2.

[fxia22]

It seems that you have multiple versions of cuda installed in /usr/local, you can expose the version used in nvidia-smi by soft linking cuda to it. For example:

[rainprob]

@fxia22
I tried to download cuda11.2 for runtime that has the same version as cuda driver version. But the results shown in web_ui is still the same.

[fxia22]

@rainprob did you build the code again? You would need to run ./build.sh build_local in GibsonEnv folder.

[rainprob]

@fxia22 yes, but the buiding result shows some warnings. I am not sure if the warnings mean building failure and will lead a wrong camera output.

[rainprob]

@fxia22

[lhy0807]

same here