HorizonStream: Long-Horizon Attention for Streaming 3D Reconstruction

Researchers introduced HorizonStream, a Transformer for streaming 3D reconstruction built with long-range and short-range geometric attention.

The model uses Metric Readout Tokens to recover pose and scale, helping stabilize online reconstruction over time.

It generalizes from short training clips to sequences longer than 10,000 frames, with bounded memory and strong accuracy.

The approach targets drift and instability in long-horizon 3D reconstruction, making it promising for robotics, mapping, and AR/VR.