RAIL-BENCH Challenges
RAIL-BENCH comprises five benchmark challenges for camera-based perception in the railway domain. Each challenge provides a curated real-world dataset, a defined evaluation protocol and a public leaderboard on Codabench.
New dataset release
RAIL-BENCH Rail
Rail Track Detection
Detect all visible rails as polylines, the basis for determining the region of interest for obstacle detection.
2,500 imagesChamferAP & LineAP
RAIL-BENCH Object
Object Detection
Detect and classify trains, signals, signal and catenary poles, road vehicles, bicycles and persons.
2,500 imagesmAP (COCO)
RAIL-BENCH Vegetation
Vegetation Segmentation
Segment low- and high-growing vegetation that may intrude into the clearance gauge of the track.
740 imagesmIoU
RAIL-BENCH Tracking
Multiple Object Tracking
Track all persons across consecutive video frames while the train passes by crowded railway platforms.
4 video sequencesHOTA
RAIL-BENCH Odometry
Monocular Visual Odometry
Estimate the motion of the train from a single front-facing camera, independent of GPS and wheel odometry.
50 sequences, 11.7 kmATE & RTE