Temporally Consistent Horizon Lines

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}
}

Setup

Get the code:

git clone --recurse-submodules https://github.com/fkluger/tchl.git
cd tchl
git submodule update --init --recursive

Set up the Python environment using Anaconda:

conda env create -f environment.yml
source activate tchl
export PYTHONPATH=./

Download the preprocessed KITTI Horizon data or generate it yourself.

Pre-trained Models

You can download the pre-trained model weights here:

• Temporally consistent ConvLSTM CNN
• Single-frame baseline
• Non-temporal (ablation study)
• Naïve residual (ablation study)

Training

In order to train the temporally consistent ConvLSTM network on KITTI Horizon, simply run:

python convlstm_net/train.py --convlstm --skip --max_error_loss --dataset_path PATH_TO_PREPROCESSED_DATASET 

For the single frame baseline, run:

python convlstm_net/train.py --seqlength 1 --batch 128 --max_error_loss --dataset_path PATH_TO_PREPROCESSED_DATASET 

Evaluation

In order to evaluate the temporally consistent CNN on KITTI Horizon, run:

python convlstm_net/evaluate.py --whole --skip --convlstm --cpu --load temporally_consistent.ckpt --set test --dataset_path PATH_TO_PREPROCESSED_DATASET

For the single-frame baseline, run:

python convlstm_net/evaluate.py --whole --cpu --load single_frame.ckpt --set test --dataset_path PATH_TO_PREPROCESSED_DATASET