HorizonStream: Long-Horizon Attention for Streaming 3D Reconstruction
Researchers have introduced HorizonStream, a novel Transformer-based architecture designed for long-horizon attention in streaming 3D reconstruction. This method addresses limitations in existing approaches that struggle with drift and jitter over extended sequences by explicitly factorizing geometric propagation as an evidence influence kernel. HorizonStream utilizes Geometric Linear Attention for multi-timescale evidence propagation and Geometric Local Attention with Spatiotemporal RoPE for reliable 3D matching, enabling stable reconstruction of sequences over 10,000 frames with constant memory and linear time. AI
IMPACT Advances streaming 3D reconstruction capabilities, potentially improving applications in robotics and augmented reality.