Sorry for the delay with this PR, I had some personal issue to deal with last week.
Anyway, main point of this PR is to add an interactive visualization using rerun. See this twitter thread for an example what can be done with it.
I made some other changes mainly related to packaging and dependencies:
Instead of installing the subdirectories in the slahmr folder, a slahmr package is installed now, and to account for this the imports will now have slahmr. prepended
Cleaned up the env.yaml file a bit; it is now more explicit which package comes from which channel and I've added nvcc which is needed to compile some of the dependencies from source
Replaced the deprecated / removed sklearn's linear_assignment with scipy's linear_sum_assignment folllowing https://stackoverflow.com/a/57992848
I'm quite sure these are good changes, but let me know if you would prefer these changes to be separate PRs or if you prefer to keep the current state.
The rerun vis currently uses PyTorch3D to compute vertex normals, I saw you removed PyTorch3D as a dependency, I guess to avoid installation issues. For me installing PyTorch3D from source via pip instead of via conda typically works quite reliably (see current env.yaml). I can also replace it with another alternative if you prefer not to have PyTorch3D.
I think currently (also in main) there is an issue with install.sh. The conda environment used in that script does not include cuda-toolkit / nvcc, so it relies on system-wide cuda being available. I think it would be better if the install.sh would just use conda env create -f env.yaml otherwise there are two slightly different setups right now.
Might also be good to update the env_build.yaml before merging.
Hi,
Sorry for the delay with this PR, I had some personal issue to deal with last week.
Anyway, main point of this PR is to add an interactive visualization using rerun. See this twitter thread for an example what can be done with it.
I made some other changes mainly related to packaging and dependencies:
slahmrfolder, a slahmr package is installed now, and to account for this the imports will now haveslahmr.prependedenv.yamlfile a bit; it is now more explicit which package comes from which channel and I've addednvccwhich is needed to compile some of the dependencies from sourcelinear_assignmentwith scipy'slinear_sum_assignmentfolllowing https://stackoverflow.com/a/57992848I'm quite sure these are good changes, but let me know if you would prefer these changes to be separate PRs or if you prefer to keep the current state.
The rerun vis currently uses PyTorch3D to compute vertex normals, I saw you removed PyTorch3D as a dependency, I guess to avoid installation issues. For me installing PyTorch3D from source via pip instead of via conda typically works quite reliably (see current env.yaml). I can also replace it with another alternative if you prefer not to have PyTorch3D.
I think currently (also in main) there is an issue with
install.sh. The conda environment used in that script does not include cuda-toolkit / nvcc, so it relies on system-wide cuda being available. I think it would be better if theinstall.shwould just useconda env create -f env.yamlotherwise there are two slightly different setups right now.Might also be good to update the
env_build.yamlbefore merging.