Code has been tested with Ubuntu 20.04, GCC 9.4.0, Python 3.8.18, PyTorch 1.9.1, CUDA 11.1 and RTX 3090.
First, it is recommended to create a new environment and install PyTorch and torchvision. Next, please use the following command for installation.
pip install -r requirements.txt
# install spconv
pip install spconv-cu113
python setup.py develop
# if you meet some pakage not matched errors, just pip install them individually before install pcdet
# rotate iou ops
cd pcdet/ops/rotated_iou/cuda_op
python setup.py install
# PointNet++
pip install "git+https://github.com/erikwijmans/Pointnet2_PyTorch.git#egg=pointnet2_ops&subdirectory=pointnet2_ops_lib"
In addition, you need to install CLIP and Minkowski Engine according to CLIP and Minkowski Engine.
We use ScanNetV2 in this work. You can download the processed data from CAGroup3D repo. After you download the dataset, remember to modify the DATA_PATH in tools/cfgs/dataset_configs/scannet_dataset.yaml.
| Task | Dataset | [email protected] | [email protected] | Download |
|---|---|---|---|---|
| Detection | ScanNetV2 | 73.7 | 61.1 | here |
Training on ScanNetV2, and we set num_gpus x batch_size to 4x4, run:
cd scripts
bash train.sh
Testing on ScanNetV2, run:
cd scripts
bash test.sh
The code for this task is built upon CAGroup3D.