Skip to content

Latest commit

 

History

History
75 lines (58 loc) · 2.5 KB

README.md

File metadata and controls

75 lines (58 loc) · 2.5 KB
Raw

Data Pipeline for LiDAR-RCNN in WOD

To decode points clouds and ground truth from Waymo Open Dataset.

cd tools/data_processer/
CUDA_VISIBLE_DEVICES='' python3 tfrecord_parser.py --data_folder /your/path/to/tfrecord --output_folder /your/path/to/save/processed_data --process 20
# CUDA_VISIBLE_DEVICES='' to disable GPU.

Then we get

ROOT_DIRECTORY
├── gt: 
      groud truth info.
├── pc:
      raw point cloud data.

After that, you can follow the Tutorial in mmdet3D to get proposals for LiDAR R-CNN. Briefly, we can get proposals in validation set by:

./tools/dist_test.sh configs/pointpillars/hv_pointpillars_secfpn_sbn_2x16_2x_waymo-3d-car.py checkpoints/hv_pointpillars_secfpn_sbn_2x16_2x_waymo-3d-car-9fa20624.pth 8 --out results/waymo-pp-car/results_eval.pkl \
    --format-only --options 'pklfile_prefix=results/waymo-pp-car/kitti_results' \
    'submission_prefix=results/waymo-pp-car/kitti_results'

Regarding training set, we provide some tricky solutions. Maybe there are more reasonable way.

Firstly, we modify the codes in mmdet3D, and run above inference command.

./configs/_base_/datasets/waymoD5-3d-car.py
Line 107 - ann_file=data_root + 'waymo_infos_val.pkl',
Line 107 + ann_file=data_root + 'waymo_infos_train.pkl',

./mmdet3d/core/evaluation/waymo_utils/prediction_kitti_to_waymo.py
Line 179 - filename = f'{self.prefix}{file_idx:03d}{frame_num:03d}'
Line 179 + filename = f'{file_idx}{frame_num:03d}'

./mmdet3d/datasets/waymo_dataset.py
Line 188 - waymo_tfrecords_dir = osp.join(waymo_root, 'validation')
Line 189 - prefix = '1'
Line 188 + waymo_tfrecords_dir = osp.join(waymo_root, 'training')
Line 189 + prefix = '0'

Subsequently, set your data path in config and run data_processer.py 

num_process: 10
pc_path: /your/path/to/raw point cloud
data_path: /your/mmdet3d/waymo-car/kitti_results_train.bin
gt_path: /your/path/to/groud truth info.
target_path: datasets/mmdet3d_pp
mode: train
expand_proposal_meter: 3
CUDA_VISIBLE_DEVICES='' python3 data_processer.py config/mmdet3d_pp_train.yaml
CUDA_VISIBLE_DEVICES='' python3 data_processer.py config/mmdet3d_pp_val.yaml

After the data conversion, the folder structure should be organized as below.

target_path
├── train.rec
├── train.idx
├── val.rec
├── val.idx