MultiviewPointpillars

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.

Dependencices

1. Python 3.5+
2. pytorch_scatter
3. Pytorch (test on 1.4)
4. OpenPCDet(spconv will not be used)
5. mmdet3d(if use dynamic voxelization)

Installation

1. clone this repository

git clone git@github.com:ge25nab/MultiviewPointpillars.git

2. install the dependet libraries

pip install -r requirements.txt 

3. install the library

python setup.py develop

Data Preparation

Follow the guidence from OpenPCDet. Only support KITTI dataset so far.

Train and Evaluation

Train Multiview Pointpillars

cd ./MultiviewPointpillars
python tools/train.py --cfg_file tools/cfgs/kitti_models/multiview_2conv.yaml

Evaluate Multiview Pointpillars

cd ./MultiviewPointpillars
python tools/test.py --cfg_file tools/cfgs/kitti_models/multiview_2conv.yaml --ckpt /path/to/your/weight

Pretrained Weight

To be uploaded