RAIL-BENCH: The first perception benchmark suite for the railway domain

The development of scene understanding algorithms requires high-quality, annotated datasets as a fundamental prerequisite. These datasets must capture the full diversity of real-world railway environments - varying lighting conditions, weather, track layouts, and traffic scenarios - in order to train and evaluate models effectively. Furthermore, predefined evaluation protocols and fixed training/test splits are necessary for a standardized comparison of algorithms. To meet these requirements, we have developed RAIL-BENCH, the first perception benchmark suite for the railway domain.

RAIL-BENCH comprises five challenges: rail track detection, object detection, vegetation segmentation, multi-object tracking, and monocular visual odometry. It provides curated datasets drawn from diverse real-world scenarios, along with standardized evaluation metrics and public leaderboards.

• Visit our RAIL-BENCH website here.
• Contact: M.Sc. Annika Bätz

RAIL-BENCH Challenges:

	RAIL-BENCH Rail Task: Detection of all visible rails in the form of polylines.
	RAIL-BENCH Object Task: Detection of dynamic objects (trains, persons, road vehicles, bicycles) and infrastructure elements (signals, signal poles, catenary poles), which are marked with bounding boxes.
	RAIL-BENCH Vegetation Task: recognition and classification of vegetation into low and high growing type.
	RAIL-BENCH Tracking Task: detect and track pedestrians across frames within a short video sequence.
	RAIL-BENCH Odometry Task: estimation of the ego motion of the camera mounted on the train.

Our project is funded by the innovation initiative mFUND by the Federal Ministry of Transport
        

and supported by DB InfraGO, Digitale Schiene Deutschland, and German Centre for Rail Traffic Research.