This is a tool for creating 3D instance segmentation annotations for the KITTI object detection dataset. The code loads in the KITTI bounding box object annotations and gives points initial labels based on whether they fall within a ground truth bounding box. The user then cleans up the annotations using a point cloud viewer GUI.

The Point Processing Toolkit is used for visualization and annotation.

Install requirements with pip

pip install -r requirements.txt

Download the KITTI object detection dataset, or if you already have it downloaded, modify the paths in config.py to point to the locations of the directories for Velodyne data, ground truth labels, calibration, and saved segmentation annotation outputs.

Default folder structure looks like this:

PROJECT_ROOT/
  - kitti/
    - data_object/
      - training/
        - calib/
        - image_2/
        - label_2/
        - velodyne/
        - gt_segmentation/  <- The annotator will create this directory. Output segmentations go here.

So, if for example you try to annotate frame number 507, the program will look for the Velodyne point cloud at kitti/data_object/training/velodyne/000507.bin.

So far, this code has only been run on Mac OS X, however pptk can also be installed on Windows and Linux so it should also be possible to run the annotator on those operating systems.

To start the annotator, just run

python main.py

Enter the frame number to annotate, and the point cloud viewer will appear. Instructions will appear in the Python terminal window - keep the terminal in view while the point cloud viewer is open.

You can also pass in as input a list of frame numbers to annotate:

python main.py -f order_file.txt

Where order_file is a list of frame numbers (valid frame numbers in the KITTI training set are 0 to 7480) separated by line breaks.

• Some frames include objects with class 'DontCare' - these can be skipped (just press ENTER).