RailLabel Providerkit

A library for annotation providers of raillabel-formatted data.

Documentation

Read the full documentation on Github pages.

Installation

You can install the latest released version directly from PyPI.

pip install raillabel-providerkit

For users who prefer a graphical interface (Mac and Windows):

pip install raillabel-providerkit[gui]

To set up a development environment, clone the project and install it into a virtual environment.

# Note that contributors should clone via ssh instead as GitHub does not allow pushing via https!
git clone https://github.com/dbinfrago/raillabel-providerkit
cd raillabel-providerkit
python -m venv .venv

source .venv/bin/activate.sh  # for Linux / Mac
.venv\Scripts\activate  # for Windows

pip install -U pip pre-commit
pip install -e '.[docs,test,gui]'
pre-commit install

Usage

GUI Application (Recommended for Non-CLI Users)

For users who prefer a graphical interface instead of the command line, launch the GUI with:

python -m raillabel_providerkit.gui

Or from Python:

from raillabel_providerkit.gui import launch_gui
launch_gui()

The GUI provides:

• 📁 Easy folder selection for input scenes and output results
• 🎯 Built-in ontology selection (OSDAR26, AutomatedTrain, OSDAR23)
• 📊 Real-time progress tracking
• ✅ Visual validation results

Requirements: Install with pip install raillabel-providerkit[gui]

Command Line Usage

You can use the validation functionality directly from the command line. To get a list of supported arguments, use this command:

python -m raillabel_providerkit --help

To simply validate all scenes in a folder and output any detected issues to an output folder (in .json format), use this command:

python -m raillabel_providerkit /path/to/folder_containing_scenes/ /path/to/output_folder

If you want to output in .csv format instead of .json format, you can use this command:

python -m raillabel_providerkit /path/to/folder_containing_scenes/ /path/to/output_folder --use-csv --no-json

Horizon Validation Tolerance

For scenes where the horizon calculation might be slightly off due to calibration inaccuracies, you can add a tolerance buffer. This only applies to track and transition annotations:

# Use default 10% tolerance above the calculated horizon line
python -m raillabel_providerkit validate /path/to/scenes/ /path/to/output/

# Or customize the tolerance
python -m raillabel_providerkit validate /path/to/scenes/ /path/to/output/ --horizon-tolerance 15.0

# With ontology validation
python -m raillabel_providerkit validate /path/to/scenes/ /path/to/output/ --ontology config/ontologies/osdar26.yaml --horizon-tolerance 10.0

The tolerance is specified as a percentage. Annotations within this buffer zone above the horizon are considered valid. Default is 10.0% to account for typical calibration variations. This helps reduce false positives when horizon calculations are slightly inaccurate. Note: Horizon validation only checks track and transition annotations, not other object types.

Exporting Scenes

Export multiple scenes to different formats (JSON, CSV) for analysis or integration with other tools:

# Export to JSON format
python -m raillabel_providerkit export /path/to/scenes/ /path/to/output/ --format json

# Export to CSV format (creates separate files for annotations, metadata, objects, sensors)
python -m raillabel_providerkit export /path/to/scenes/ /path/to/output/ --format csv

# Export to both formats
python -m raillabel_providerkit export /path/to/scenes/ /path/to/output/ --format json --format csv

Or programmatically from Python:

from pathlib import Path
from raillabel_providerkit import export_scenes

stats = export_scenes(
    input_folder=Path("/path/to/scenes/"),
    output_folder=Path("/path/to/output/"),
    formats=["json", "csv"]
)
print(f"Exported {stats['exported']} scenes, {stats['errors']} errors")

Using Built-in Ontologies

RailLabel Providerkit comes with pre-built ontology parameter files that can be used directly:

# Using OSDAR26 ontology
python -m raillabel_providerkit /path/to/scenes/ /path/to/output/ --ontology config/ontologies/osdar26.yaml

# Using AutomatedTrain ontology
python -m raillabel_providerkit /path/to/scenes/ /path/to/output/ --ontology config/ontologies/automatedtrain.yaml

Or programmatically from Python:

from raillabel_providerkit import validate, get_ontology_path

# Using a built-in ontology
ontology_path = get_ontology_path("osdar26")
issues = validate("path/to/scene.json", ontology=ontology_path)

# List available ontologies
from raillabel_providerkit import list_available_ontologies
print(list_available_ontologies())  # ['osdar23', 'osdar26', 'automatedtrain']

You can also provide a custom ontology file:

python -m raillabel_providerkit /path/to/folder_containing_scenes/ /path/to/output_folder --ontology /path/to/custom-ontology.yaml

Supported Ontologies / Parameter Files

Pre-built ontology parameter files are provided and accessible via the API:

Ontology	Name	Dataset	Description
osdar23	OSDAR23	OSDAR23	Original railway environment annotation ontology. Includes standard occlusion ranges (0-25%, 25-50%, 50-75%, 75-99%, 100%) and core railway classes.
osdar26	OSDAR26	OSDAR26	Extended railway environment ontology with 25 object classes. Features comprehensive signal aspects (Hp, Ks, Vr, Zs, Sh variants), updated occlusion ranges (0-24%, 25-49%, 50-74%, 75-99%, 100%), and additional classes.
automatedtrain	AutomatedTrain	AutomatedTrain	Specialized ontology for automated train perception and safety-critical annotation. Includes obstacle detection, platform recognition, level crossings, speed signs.

OSDAR26 Classes (25 total)

• Persons: person, crowd
• Personal Mobility: personal_item, pram, scooter, wheelchair
• Vehicles: bicycle, group_of_bicycles, motorcycle, road_vehicle
• Animals: animal, group_of_animals
• Railway Vehicles: train, wagon, drag_shoe
• Track Infrastructure: track, transition, switch
• Signaling: signal, signal_pole, signal_bridge, catenary_pole, buffer_stop
• Hazards: flame, smoke

AutomatedTrain Classes

The AutomatedTrain ontology supports safety-critical classes for automated train operation:

• Core Classes: person, crowd, train, wagon, track, transition, switch
• Signaling: signal, signal_pole, signal_bridge, catenary_pole, buffer_stop
• Vehicles: road_vehicle, bicycle, motorcycle, animal
• Safety-Critical: obstacle, platform, level_crossing, speed_sign

Supported Sensors

RailLabel Providerkit supports sensors from both OSDAR23 and OSDAR26 datasets. The sensor validation checks that all sensors in a scene have recognized names and correct types.

OSDAR23 Sensors

Sensor ID	Type	Description
rgb_center, rgb_left, rgb_right	Camera	Standard RGB cameras
rgb_highres_center, rgb_highres_left, rgb_highres_right	Camera	High-resolution RGB cameras
rgb_longrange_center, rgb_longrange_left, rgb_longrange_right	Camera	Long-range RGB cameras
ir_center, ir_left, ir_right	Camera	Infrared cameras
lidar	Lidar	Single lidar sensor
radar	Radar	Single radar sensor
gps_imu	GpsImu	GPS/IMU sensor

OSDAR26 Sensors

Sensor ID	Type	Description
rgb_12mp_left, rgb_12mp_middle, rgb_12mp_right	Camera	12 megapixel RGB cameras
rgb_5mp_left, rgb_5mp_middle, rgb_5mp_right	Camera	5 megapixel RGB cameras
ir_left, ir_middle, ir_right	Camera	Infrared cameras
lidar_merged	Lidar	Merged lidar data
radar_cartesian	Radar	Radar in cartesian coordinates

Adding Custom Sensors

To add support for additional sensors, edit the SENSOR_METADATA dictionary in raillabel_providerkit/_util/_sensor_metadata.py:

from raillabel.format import Camera, Lidar, Radar, GpsImu

SENSOR_METADATA = {
    # ... existing sensors ...

    # Add your custom sensors
    "my_custom_camera": Camera,
    "my_custom_lidar": Lidar,
}

Calibration Detection

The horizon validation automatically detects the calibration format (OSDAR23 vs OSDAR26) based on sensor naming patterns. OSDAR26 uses different extrinsics rotation axis conventions compared to OSDAR23. The detection is performed once per scene by checking all sensor names - if any sensor matches the OSDAR26 pattern (e.g., rgb_12mp_left, ir_middle), the OSDAR26 calibration is applied to the entire scene.

Contributing

We'd love to see your bug reports and improvement suggestions! Please take a look at our guidelines for contributors for details.

Licenses

This project is compliant with the REUSE Specification Version 3.0.

Copyright DB InfraGO AG, licensed under MIT (see full text in LICENSE)

Dual Licensing:

• Main code (MIT): Permissive open-source license allowing commercial use, modification, and distribution with minimal restrictions
• Configuration files (CC0-1.0): Public domain dedication for ontology files and configuration data
• Documentation: Available under both MIT and Apache 2.0 where applicable

See individual files for their specific licenses.