runjian-chen / co3 Goto Github PK

Hi Chen!

Thank you again for your excellent paper! I am quite curious that STRL is reported to have a much higher score than BYOL in your paper (according to Table 1 & 2) which seemed quite amazing to me at first, as STRL is basically BYOL but extended to the 3D realm with little modification.

I guess it is because BYOL is trained following the point-wise objective (akin to your Cooperative Contrastive Objective but without negatives) while STRL is trained with a global objective (i.e., predict the global feature vector for a corresponding view). It seems counter-intuitive that a finer-grained pretraining target reaps worse results.

I am not posing questions on the numbers though, I will be pretty glad anyway should you be kind enough to open-source the baseline methods that you implemented. I am more curious about the high-level implications. Does this imply that point-wise contrastive learning (aka registration) is not a good pretraining objective even compared with a global objective?

Hi Chen! Thanks a lot for your excellent work and its open-sourced code!

I am recently trying out different environment builds on a CUDA 11 machine. I found out an installation that worked for me, looks similar to your claimed environment (python 3.8 + torch 1.8), bug-free during installation, and require no source-code compiling.

A major issue with the previous version is that mmdet3d, mmdet, mmcv and mmsegmentation just released a whole bunch of incompatible version so that pip install XXX.git will not work as intended (latest incompatible version is installed).

My installation reads as follows:

conda create -n open-mmlab python=3.8 -y
conda activate open-mmlab
conda install pytorch=1.8 torchvision=0.9.0 cudatoolkit=11.1 -c pytorch -c nvidia
pip install mmcv-full==1.4.5 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.8.0/index.html --no-cache-dir
pip install mmdet==2.22.0 mmsegmentation==0.20.0
wget https://github.com/open-mmlab/mmdetection3d/archive/refs/tags/v0.18.0.tar.gz
tar -xvf v0.18.0.tar.gz
cd mmdetection3d
pip install -v -e .
pip install git+https://github.com/klintan/pypcd.git
pip install timm