Rethinking the State Update Gate for Long-Sequence Recurrent 3D Reconstruction

Streaming 3D reconstruction suffers from long-sequence drift because existing per-token gates are structurally bounded, limiting memory to about three frames. The authors introduced Adaptive Frame Gating (AFG), a parameter-free scalar gate that scales the update strength based on scene novelty. This extends the effective memory horizon from three to up to 64 frames, significantly improving performance in long sequences.

Open the full episode

More from AI Radio