This repository is an implementation of PointPainting on the Waymo Open Dataset.
Download v1.4.1 of the Waymo Open Dataset following the instruction. Format the segments under the following folder structure.
waymo
|- waymo_format
|- training
|- validation
|- testing
Create a conda environment
conda create -n pp python=3.10
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
conda install tqdm
conda install numba==0.57
pip install opencv-python
pip install open3d
pip install tensorboard
pip install waymo-open-dataset-tf-2-11-0==1.6.1
conda activate pp
cd ops
python setup.py develop
First convert the dataset to the KITTI format. This will create a kitti_format folder under your waymo directory.
cd data_prep
python convert_data.py --waymo_root [path/to/waymo]
Next paint the lidar points using a trained segmentation model.
cd painting
python painting.py --training_path [path/to/waymo]/kitti_format/training/ --model_path [path/to/segmentation/model]
Create the info file used for training
cd data_prep
python create_info.py --waymo_root [path/to/waymo] --painted --create_info
To train on painted lidar points.
conda activate pp
torchrun --nproc_per_node=[gpus] train.py --data_root [path/to/waymo]/kitti_format/ --painted --cam_sync --save-path [checkpoint/path] --max_epoch [num of epochs] --ckpt_freq_epoch [freq]
To evaluate the mAP.
conda activate pp
python evaluate.py --ckpt [checkpoint/path] --data_root [path/to/waymo]/kitti_format/ --painted --cam_sync
To perform inference
conda activate pp
python inference.py --data_root [path/to/waymo]/kitti_format --lidar_detector [pointpillars/checkpoint/path] --segmentor [deeplab/checkpoint/path] --painted --cam_sync
This repository makes use of the open source from PointPillars, MMDet3D, DeepLabV3Plus-Pytorch, and PointPainting.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent