If you use this code, please cite our paper:

@inproceedings{kluger2020temporally,
  title={Temporally Consistent Horizon Lines},
  author={Kluger, Florian and Ackermann, Hanno and Yang, Michael Ying and Rosenhahn, Bodo},
  booktitle={2020 International Conference on Robotics and Automation (ICRA)},
  year={2020}
}

Prerequisites

Get the code:

git clone https://github.com/fkluger/kitti_horizon.git
cd kitti_horizon

Install required libraries, e.g. via pip or conda:

• pykitti, numpy, scikit-image, opencv
• optional: pytorch, Pillow

Download the KITTI raw dataset.

Usage

kitti_horizon_raw.py

Contains the KITTIHorizonRaw class, which extracts horizon lines in image coordinates from the KITTI raw data. It also applies image scaling and padding in order to achieve consistent image resolutions. See python kitti_horizon_raw.py --help for a demo.

process_kitti_horizon_raw.py

Pre-processes the KITTI Horizon dataset (using the KITTIHorizonRaw class) and stores it as pickle files, in order to speed up the training process. See python process_kitti_horizon_raw.py --help for usage.

kitti_horizon_torch.py

Provides the KITTIHorizon class, derived from the PyTorch Dataset class. Requires the pre-processed KITTI Horizon dataset. For usage as a training/validation dataset in PyTorch. See python kitti_horizon_torch.py --help for a demo.