This page presents LARD V2, the second version of our dataset and generator. LARD V1 can be found here.

Landing Approach Runway Detection (LARD) is a dataset of aerial front view images of runways designed for aircraft landing phase. This second version contains over 100 000 synthetic images of runways from the ~250 airports with the most traffic worldwide.

LARD is also a synthetic image generator which allows to generate images from 5 different data sources: Google Earth, Bing Maps, ArcGIS, XPlane and Flight Simulator. In this second version of LARD, we provide an interface for generating landing scenarios based on Streamlit.

LARD V1	Additional data sources in LARD V2
Google Earth	(Top) left: XPlane, right: Flight Simulator
	(Bottom) left: Bing Maps, right: ArcGIS

• Download LARD V2 dataset
• Launch Streamlit Application
• [Read our paper] (To be published)

• Streamlit Interface
• LARD V2 dataset
• Image Generation Recommendations
• Models
• See Also
• Authors
• Citation
• License

LARD V2 now provides a user interface to simplify the generation of new data, based on Streamlit.

• pip install: We recommend using the requirements.txt to install packages in your environment with:

pip install -r requirements.txt

You can run the Streamlit interface with:

streamlit run ./Home_page.py

This will open a new window in your browser, which will guide you through the process of generating new scenarios. This interface provides the same functionalities as the usual notebooks 01_scenario_generation.ipynb and 02_labeling.ipynb, but in a more convenient way.

If needed, these notebooks are still available if needed, please refer to the section 2 of the LARD V1 Readme for more details.

• LARD V2 - Download

This dataset is dedicated to the detection of a runway from an aircraft during approach and landing phases. In our work, we only consider the approach segment which ranges from -6000 m to -280 m along-track distance from Landing Threshold Point (LTP). Throughout this segment, the acceptable aircraft positions remain similar but not equivalent to the ODD of LARD V1: the lateral path angle lies within [-3°,3°], originating from the LTP, while the vertical path angle takes its values in [-1.8°,-5.2°] w.r.t. the Vertical Reference Point (VRP), a point on the centerline located 305 m beyond the LTP.

For the aircraft attitude, the range for the pitch stays constant, between [-15°, +5°], which translates to an average nose-down orientation of about 5° relative to the horizon. However, to define appropriate ranges for the aircraft roll and yaw, the approach cone is split into 3 segments as follows:

The first 3 parameters correspond to the position of the aircraft relative to the runway, while the last 3 parameters correspond to its orientation. Please not that these

The dataset available on HugginFace is structured as follows:

A folder images contains all images from LARD V2. It is further partitioned by data sources, which in turn contain one folder for each airport:

    images/
    | arcgis/
    | | CYEG/
    | | | CYEG-2-20-12-30__10-smpl__10h40_000_ArcGIS.jpg
    | | | CYEG-2-20-12-30__10-smpl__10h40_001_ArcGIS.jpg
    | | | ...
    | | CYVR/
    | | | ...
    | bingmaps/
    | | CYEG/
    | | | CYEG-2-20-12-30__10-smpl__10h40_000_BingMaps.jpg
    | | | CYEG-2-20-12-30__10-smpl__10h40_001_BingMaps.jpg
    | | | ...
    | | CYVR/
    | | | ...

The labels of the dataset are provided in a series of .csv files next to the images/ folder.

The dataset is already splitted between train and test. This was performed using the list of airports considered, with approximately 50% or airports in the training set, and 50% in the test set, selected at random.

The labels of the dataset are provided in a series of .csv next to the images/ folder. Each .csv file corresponds to a single data source, and is dedicated to either training or testing as mentioned in the file name.

• First, a few general information about the pictures are provided, such as the relative path of the picture or its width and height. The images all come in the same 1024×1024 resolution.

• Then, data about the origin of the picture are given, such as the initial scenario which generated the image, or the type of data source that produced the data.

• The next columns indicate the airport and runway targeted during landing.

• Additional columns provide information about the pose, its coordinates, several parameters that provide information about current pose relative to some runway keypoints, and the attitude of the aircraft (yaw, pitch, roll)

• A specific column inform if the corresponding runway visible on the image is considered IN_ODD or IN_EXTENDED_ODD. Please refer to our paper for more details about this concept.

• Finally, the last columns provide the pixel coordinates of each corner of the runways visible on the picture.

This section describes the different interfaces used to generate the images expected to be labeled by the tool. The production of image based on Google Earth Studio is still support (as in LARD V1), but requires to generate the .esp file along the .yaml one during the scenario generation step (with Streamlit or with the dedicated notebooks). All other sources of data take the .yaml file as input and outputs the expected images.

The current streamlit interface allows to generate .esp files that can be used in Google Earth Studio as described in LARD V1. Please refer to the section 2-bis of the LARD V1 Readme for more details about this functionality.

You can generate images using XPlane 12. It requires to possess a commercial version of XPlane 12, and to follow everything explained directly in this associated github repository:

(In Construction)

You can generate images using Flight simulator. It requires to possess Flight Simulator 2020, and to follow everything explained directly in this associated github repository: github.com/JeanBriceGinestet/GeoFlight-Replay.

(In Consolidation)

The resulting images should be made available in a "footage" folder, with the associated yaml files.

The data sources ArcGIS, Bing Maps and Google Earth will be made available through a dedicated interface based on Cesium.

(In Construction)

Our latest paper on LARD V2 compares models trained on multiple configurations of LARD V2 data. For reproductibility, these models are available on a separate public repository with a AGPL license.

LARD_V2 models: https://github.com/deel-ai-papers/Yolo_models_LARD_V2

More from the DEEL project:

• Xplique a Python library exclusively dedicated to explaining neural networks.
• deel-lip a Python library for training k-Lipschitz neural networks on TF.
• Influenciae Python toolkit dedicated to computing influence values for the discovery of potentially problematic samples in a dataset.
• deel-torchlip a Python library for training k-Lipschitz neural networks on PyTorch.
• DEEL White paper a summary of the DEEL team on the challenges of certifiable AI and the role of data quality, representativity and explainability for this purpose.

        

You can read our first paper at https://hal.science/hal-04056760.

If you use LARD as part of your workflow in a scientific publication, please consider citing the following paper:

@unpublished{ducoffe:hal-04056760,
  TITLE = {{LARD - Landing Approach Runway Detection - Dataset for Vision Based Landing}},
  AUTHOR = {Ducoffe, M{\'e}lanie and Carrere, Maxime and F{\'e}liers, L{\'e}o and Gauffriau, Adrien and Mussot, Vincent and Pagetti, Claire and Sammour, Thierry},
  URL = {https://hal.science/hal-04056760},
  NOTE = {working paper or preprint},
  YEAR = {2023},
  MONTH = Apr,
  KEYWORDS = {open-source dataset ; vision based landing ; ML},
  PDF = {https://hal.science/hal-04056760/file/main.pdf},
  HAL_ID = {hal-04056760},
  HAL_VERSION = {v1},
}

The package is released under MIT license.