Synthetic data · Real footage
Auto-labeled synthetic obstacles on real railway footage
Synthetic obstacles composited onto real cab-view footage: rail segmentation per frame, world-locked placement scaled against the track gauge, and YOLO labels that come out of the placement step. This run: 960 frames, 8 classes, 3,145 annotations.
SYNTHETIC COMPOSITE
Synthetic obstacles were composited onto real cab-view footage; no real event is shown. The boxes are ground truth from the placement step, not model predictions.
01What the film shows
Four views run on the same frame. Rail segmentation solves the ego path per frame, with the track gauge measured in pixels, and everything downstream is anchored to that path. Obstacles are placed world-locked, not painted onto the screen: they stay put in the scene as the camera moves, and they are scaled against the rail gauge so a 40 cm object is 40 cm at any distance.
The boxes are not predictions. The position, scale and class of every object are known when it is placed, so the YOLO label file falls out with the frame, with no frame-by-frame annotation. Synthetic data does not replace field data; it fills the part of the distribution the field will not give you.
