KITTI Object data transformation and visualization

Dataset

Download the data (calib, image_2, label_2, velodyne) from Kitti Object Detection Dataset and place it in your data folder at kitti/object

The folder structure is as following:

kitti
    object
        testing
            calib
               000000.txt
            image_2
               000000.png
            label_2
               000000.txt
            velodyne
               000000.bin
            pred
               000000.txt
        training
            calib
               000000.txt
            image_2
               000000.png
            label_2
               000000.txt
            velodyne
               000000.bin
            pred
               000000.txt

Install locally on a Ubuntu 16.04 PC with GUI

• start from a new conda enviornment:

(base)$ conda create -n kitti_vis python=3.7 # vtk does not support python 3.8
(base)$ conda activate kitti_vis

• opencv, pillow, scipy, matplotlib

(kitti_vis)$ pip install opencv-python pillow scipy matplotlib

• install mayavi from conda-forge, this installs vtk and pyqt5 automatically

(kitti_vis)$ conda install mayavi -c conda-forge

• test installation

(kitti_vis)$ python kitti_object.py --show_lidar_with_depth --img_fov --const_box --vis

Note: the above installation has been tested not work on MacOS.

Install remotely

Please refer to the jupyter folder for installing on a remote server and visulizing in Jupyter Notebook.

Visualization

1. 3D boxes on LiDar point cloud in volumetric mode
2. 2D and 3D boxes on Camera image
3. 2D boxes on LiDar Birdview
4. LiDar data on Camera image

$ python kitti_object.py --help
usage: kitti_object.py [-h] [-d N] [-i N] [-p] [-s] [-l N] [-e N] [-r N]
                       [--gen_depth] [--vis] [--depth] [--img_fov]
                       [--const_box] [--save_depth] [--pc_label]
                       [--show_lidar_on_image] [--show_lidar_with_depth]
                       [--show_image_with_boxes]
                       [--show_lidar_topview_with_boxes]

KIITI Object Visualization

optional arguments:
  -h, --help            show this help message and exit
  -d N, --dir N         input (default: data/object)
  -i N, --ind N         input (default: data/object)
  -p, --pred            show predict results
  -s, --stat            stat the w/h/l of point cloud in gt bbox
  -l N, --lidar N       velodyne dir (default: velodyne)
  -e N, --depthdir N    depth dir (default: depth)
  -r N, --preddir N     predicted boxes (default: pred)
  --gen_depth           generate depth
  --vis                 show images
  --depth               load depth
  --img_fov             front view mapping
  --const_box           constraint box
  --save_depth          save depth into file
  --pc_label            5-verctor lidar, pc with label
  --show_lidar_on_image
                        project lidar on image
  --show_lidar_with_depth
                        --show_lidar, depth is supported
  --show_image_with_boxes
                        show lidar
  --show_lidar_topview_with_boxes
                        show lidar topview
  --split               use training split or testing split (default: training)

$ python kitti_object.py

Specific your own folder,

$ python kitti_object.py -d /path/to/kitti/object

Show LiDAR only

$ python kitti_object.py --show_lidar_with_depth --img_fov --const_box --vis

Show LiDAR and image

$ python kitti_object.py --show_lidar_with_depth --img_fov --const_box --vis --show_image_with_boxes

Show LiDAR and image with specific index

$ python kitti_object.py --show_lidar_with_depth --img_fov --const_box --vis --show_image_with_boxes --ind 1 

Show LiDAR with label (5 vector)

$ python kitti_object.py --show_lidar_with_depth --img_fov --const_box --vis --pc_label

Demo

2D, 3D boxes and LiDar data on Camera image

Credit: @yuanzhenxun

LiDar birdview and point cloud (3D)

Show Predicted Results

Firstly, map KITTI official formated results into data directory

./map_pred.sh /path/to/results

python kitti_object.py -p --vis

Acknowlegement

Code is mainly from f-pointnet and MV3D