Synthetic data · Railway obstacle detection
Synthetic data for railway obstacle detection: the rails are the ruler
You cannot wait for an obstacle on the track, and you cannot stage one. So the training data is built on real cab-view video, and every step is measured: the rails give horizon and scale without calibration, a placed obstacle stays fixed in the scene, and its label follows what stays visible.
Chapters
SYNTHETIC DATA
The footage was generated in Blender and is not a field recording. The boxes are ground truth produced at render time, not model predictions. Internal R&D work.
01What the film shows
Obstacles on the track are exactly the data a detector needs and nobody can collect: they are rare, dangerous and impossible to stage. So the training data is built on real cab-view video. The ego-path model finds the train's own corridor in every frame, and the geometry does the measuring: the gap between the rails narrows in a straight line toward the horizon, so the rails give both the horizon and the scale, with no camera calibration.
Ground points tracked from frame to frame measure the distance travelled, so a placed obstacle stays fixed in the scene and grows the way a real object would. Its label is a polygon of what is actually visible, taken from the image's own outline, and it keeps following that part as the obstacle leaves the frame. The obstacle images come from a generated library that is checked automatically, and no synthetic image reaches training until a person approves it.
Fully synthetic visualization: the railway is a 3D scene and the obstacle images come from our own generated asset library. The pipeline itself works on real cab-view video, which is not shown in this film. The figures are Spikedge R&D bench measurements from September 2026: 1,672 frames of one cab-view clip, where a classical texture mask put the track corridor in the wrong place in 62.6% of frames and the ego-path model in 0%, and one validation scene with 8 obstacles. Detection performance in the field has not been measured yet.
02Transcript
Some events are too rare to record, and too dangerous to stage.
A detector still needs thousands of labelled frames of them.
We start from real cab-view video, and find the train's own path.
No calibration needed. The gap between the rails narrows in a straight line: the rails are the ruler.
Ground points measure the distance travelled, so the obstacle stays put, and grows as it should.
The label comes from geometry, and follows what stays visible.
The obstacles come from a generated library, checked automatically.
And no synthetic image reaches training until a person approves it.
We measure every step. What we haven't measured yet, we don't claim.
