Installation

The codebase is built on PyTorch 1.10, MMCV 1.x, and MMDetection 2.x. The following environment matches the versions used by the project.

1. Create the environment

conda create -n daoccv2 python=3.8 -y
conda activate daoccv2

2. Install PyTorch

Install PyTorch 1.10.2 and torchvision 0.11.3 for CUDA 11.3:

pip install torch==1.10.2+cu113 torchvision==0.11.3+cu113 -f https://download.pytorch.org/whl/torch_stable.html

If a different CUDA toolkit is used, select the matching PyTorch build and ensure it remains compatible with the MMCV build installed below.

3. Install MMCV and the Python dependencies

pip install mmcv-full==1.4.0 -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.10.0/index.html
pip install mmdet==2.20.0
pip install nuscenes-devkit==1.1.11
pip install torchpack==0.3.1
pip install numba==0.48.0
pip install numpy==1.22.0
pip install prettytable==3.11.0
pip install ninja==1.11.1.1
pip install einops==0.8.0
pip install yapf==0.40.0
pip install pillow==8.4.0
pip install Shapely==1.7.1
pip install pycocotools==2.0.7
pip install scikit-image==0.19.3
pip install open3d==0.19.0

4. Waymo development kit (optional)

This step is required when preprocessing Occ3D-Waymo. The converter currently imports TensorFlow and the Waymo package exposed by this development-kit build:

pip install waymo-open-dataset-tf-2-6-0==1.4.8

5. Install DAOccV2

git clone https://github.com/ZYang2077/DAOccV2.git
cd DAOccV2
python setup.py develop