About RAIL-BENCH
RAIL-BENCH is the first perception benchmark suite for the railway domain. Automated train operation requires robust camera-based perception, yet the railway domain lacks public benchmark suites with standardized evaluation protocols that would enable reproducible comparison of approaches. RAIL-BENCH addresses this gap by providing curated training and test datasets drawn from diverse real-world scenarios, evaluation metrics, and public scoreboards.
Five Benchmark Challenges
-
RAIL-BENCH Rail — Rail Track Detection:
Recognition of rail tracks modelled as polylines marking the left and right rail. Predictions are evaluated using the ChamferAP as well as the novel LineAP metric — a segment-based average precision that evaluates geometric accuracy of polyline predictions independently of instance-level grouping. -
RAIL-BENCH Object — Object Detection:
Recognition of objects belonging to one of seven categories: train, signal, signal pole, catenary pole, road vehicle, bicycle, and person. Objects are identified by a bounding box together with a class label. -
RAIL-BENCH Vegetation — Vegetation Segmentation:
Pixel-wise recognition of vegetation in the track area and surroundings. Encroaching vegetation poses a safety-critical hazard when violating the structural clearance gauge of the railway. The task requires classifying each pixel into low-growing vegetation (e.g., grasses and herbs) or high-growing vegetation (e.g., trees and shrubs). -
RAIL-BENCH Tracking — Multiple Object Tracking:
Accurate association of person detections across frames within short video sequences. Multi-object tracking is particularly challenging for crowds of people on platforms. -
RAIL-BENCH Odometry — Monocular Visual Odometry:
Estimation of the ego motion of the camera from monocular video sequences. Although railway vehicle motion is constrained by rails, large slip and unknown branching directions at turnouts pose considerable difficulties.
RAIL-BENCH Rail
RAIL-BENCH Object
RAIL-BENCH Vegetation
RAIL-BENCH Tracking
RAIL-BENCH Odometry
Citation
If you use Rail-Bench in your research, please cite it as follows:
A. Bätz, P. Klasek, S.-Y. Ham, P. Neumaier, M. Köppel, and M. Lauer, "Railway Artificial Intelligence Learning Benchmark (RAIL-BENCH): A Benchmark Suite for Perception in the Railway Domain," submitted to: IEEE/RSJ Int. Conf. on Intell. Robots and Syst., 2026
Terms of Use
By using RAIL-BENCH, you agree to:
- Use the datasets for research and non-commercial purposes only.
- Properly cite RAIL-BENCH in your work.
- Not redistribute or commercialize the content without permission.
Contact
For questions or inquiries, please reach out to: