HorizonStream: Long-Horizon Attention for Streaming 3D Reconstruction

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

It uses Geometric Linear Attention, Geometric Local Attention, and Metric Readout Tokens to propagate scene geometry efficiently.

Trained on short clips, the model generalizes to sequences longer than 10,000 frames while keeping memory constant and runtime linear.

The approach improves stability and reduces drift in online reconstruction, achieving state-of-the-art streaming 3D reconstruction results.