INSTALL.md

Installation

Requirements

All the codes are tested in the following environment:

• Linux (tested on Ubuntu 14.04/16.04/18.04/20.04/21.04)
• Python 3.6+
• PyTorch 1.1 or higher (tested on PyTorch 1.1, 1,3, 1,5~1.10)
• CUDA 9.0 or higher (PyTorch 1.3+ needs CUDA 9.2+)
• spconv v1.0 (commit 8da6f96) or spconv v1.2 or spconv v2.x

Install MS3D

MS3D runs on OpenPCDet pcdet v0.6.0.

a. Clone this repository.

git clone https://github.com/darrenjkt/MS3D.git

b. Using Docker

We highly recommend running this repository on docker for out-of-box usage. You can build our Dockerfile or pull our provided docker image.

docker pull darrenjkt/openpcdet:v0.6.0

For easy running of the image we provide a script. Change the file paths, number of GPUs and then run it. Use docker ps to find the container name.

bash docker/run.sh
docker exec -it ${CONTAINER_NAME} /bin/bash

c. Within the container, install pcdet and the tracker with the following commands

python setup.py develop
cd tracker && pip install -e . --user

Note that if you want to use dynamic voxelization (e.g. Voxel-RCNN), you need torch-scatter. This should already be pre-installed in our docker image but you can install with the following commands if need be.

pip install torch==1.8.1 torchvision==0.9.1
pip install torch-scatter -f https://data.pyg.org/whl/torch-1.8.1+cu111.html
python setup.py develop # rebuild repository