KITTI 3D to FAT Converter

Converts KITTI 3D Object Detection Evaluation 2017 Object Pose Estimation training dataset
to Falling Things Format.

The code was written and tested with Python 3.8.

Installation

pip3 install pyjson pyyaml numpy Pillow scipy

Prerequesites

1. Download the following from KITTI 3D Object Detection Website:

   • left color images of object data set (12 GB)
   • training labels of object data set (5 MB)
   • camera calibration matrices of object data set (16 MB)

2. Extract files so that your directory tree is as follows:

.
├── calib
│   ├── 000000.txt
│   ├── 000001.txt
│   └── ...
├── image_2
│   ├── 000000.png
│   ├── 000001.png
│   └── ...
└── label_2
    ├── 000000.txt
    ├── 000001.txt
    └── ...

Usage

usage: convert.py [-h] --kitti-dir DIR --output-dir DIR [--distance-in-cm] [--save-camera-info] [--debug]

Converts given KITTI 3D Object Detection Training Dataset to Falling Things Format.

optional arguments:
  -h, --help           show this help message and exit
  --kitti-dir DIR      Path to KITTI root directory.
  --output-dir DIR     Path to store converted dataset.
  --distance-in-cm     Distance unit for output dataset is centimeters.
  --save-camera-info   Stores CameraInfo yaml files for each frame.
  --debug              Draws cuboids on the output image.

Shortcommings

• Does not create the _object_settings.json and _camera_settings.json
• Ccclusion is handled as follows: kitti_occlusion_to_fat_occlusion = {0: 0.0, 1: 0.25, 2: 0.75, 3: 0.0}.
• Since KITTI does not provide a visibility value, the following code reproduces the visibility:

visibility = 1.0 - truncation
visibility = max(0, visibility - occlusion)