This repository contains code to reproduce the baseline results for the MLAD Challenge.
We recommend Python 3.6+ for running the code. Conda is required. The conda environment requires CUDA 10.1 to be installed. You can install the environment with:
conda env create -f environment.yml
To clone the project and install the submodules please run the following:
git clone https://github.com/pmwenzel/mlad-benchmark-baselines.git
cd mlad-benchmark-baselines
git submodule update --init
To download the pretrained models for D2-Net please run the following:
mkdir third_party/d2_net/models
wget https://dsmn.ml/files/d2-net/d2_tf.pth -O third_party/d2_net/models/d2_tf.pth
The dataset for the challenge can be downloaded here. Please make sure to download the whole dataset, including reference, training, validation, and both test sequences.
The code expects the following data structure after extraction of the files.
.
├── recording_2020-03-03_12-03-23
├── recording_2020-03-24_17-36-22
├── recording_2020-03-24_17-45-31
├── recording_2020-04-07_10-20-32
└── recording_2020-04-23_19-37-00
The baselines for our challenge are currently based on the following GitHub projects:
To run the baselines for test_sequence0 run the following:
bash run_baselines.sh /path/to/dataset /path/to/output/folder 0
To run the baselines for test_sequence1 run the following:
bash run_baselines.sh /path/to/dataset /path/to/output/folder 1
The result of each re-localization file will be saved as a .txt file in the format as expected to be submitted to the challenge.
Each line of these text-files is saved in the following way:
source_kf target_kf t_x t_y t_z q_x q_y q_z q_w
The code is provided under a BSD 3-clause license. See the LICENSE file for details. Note also the different licenses of thirdparty submodules.