Horizon Robotics has described LoFT, a model that reconstructs geometry from video, with its paper, code and weights still pending and results reported by the team.
Geometry reconstruction from video matters for autonomous driving and robotics because it turns ordinary camera footage into spatial structure - the kind of signal planners need without relying on lidar for every scene.
Doing it in video rather than single frames lets the model use temporal consistency, which reduces the flicker and scale ambiguity that plague per-frame depth estimation.
The pending release is the material caveat. Until code and weights are available, independent teams cannot verify the numbers or test the model on their own data.
For Horizon, whose business is automotive chips and software stacks, a geometry model is a component rather than a product; the strategy is to improve the perception stack its customers already license.
