MultiviewPointpillars is a variant of Pointpillars with both front view and BEV.
The code structure is based on OpenPCDet. The main idea also refers to the paper Pillar-based Object Detection for Autonomous Driving, and could be viewed as its variant re-implementation from Tensorflow to PyTorch Version.
- Python 3.5+
- pytorch_scatter
- Pytorch (test on 1.4)
- OpenPCDet(spconv will not be used)
- mmdet3d(if use dynamic voxelization)
- clone this repository
git clone git@github.com:ge25nab/MultiviewPointpillars.git
- install the dependet libraries
pip install -r requirements.txt
- install the library
python setup.py develop
Follow the guidence from OpenPCDet. Only support KITTI dataset so far.
cd ./MultiviewPointpillars
python tools/train.py --cfg_file tools/cfgs/kitti_models/multiview_2conv.yaml
cd ./MultiviewPointpillars
python tools/test.py --cfg_file tools/cfgs/kitti_models/multiview_2conv.yaml --ckpt /path/to/your/weight
To be uploaded