/Partial_Point_Clouds_Generation

Extract partial point clouds within given 3D bounding boxes from KITTI

Primary LanguagePython

Getting Started

This repo is aim to extract partial point clouds within labeled 3D bounding boxes. Currently, the KITTI dataset is supported.

Dataset Preparation

  • Please download the official KITTI 3D object detection dataset and organize the downloaded files as follows: 
    ├── kitti
│   │── ImageSets
│   │── training
│   │   ├──calib & velodyne & label_2 & image_2
│   │── testing
│   │   ├──calib & velodyne & image_2
│   
├── extract_partial_point_clouds_from_kitti.py
├── utils.py
├── visualize.py
├── calibration.py

Quick Demo

  • Here we provide a quick demo to extract partial point clouds of cars from the point clouds saved in the kitti/training/velodyne/000936.bin:

    python extract_partial_point_clouds_from_kitti.py \
    --idx 000936 \
    --category car

    Each individual 3D bounding box and the partial point clouds within it are saved in the output/ with the filename 000936_car_{point or bbox}_0.npy.

  • Visualization. Take the 000936_car_point_0.npy for example:

    python visualize.py \
    --idx 000936 \
    --category car
    --i 0

Others

The output points clouds and bounding boxes are tranformed into the coordinate system of Shapenet (x->forward, y->up) from kitti lidar (x->forward, z->up). If you prefer no such tranformation, comment the following line in the extract_partial_point_clouds_from_kitti.py:

...
box = bboxes[i]
points_canonical, box_canonical = utils.points_to_canonical(p, box)
# points_canonical, box_canonical = utils.lidar_to_shapenet(points_canonical, box_canonical)
pts_name = 'output/{}_{}_point_{}'.format(args.idx, args.category, i)
box_name = 'output/{}_{}_bbox_{}'.format(args.idx, args.category, i)
...