Behavior of MeMix on thousand-frame (kilometer-scale) sequences

Determine the behavior of MeMix, a training-free plug-in memory update module for recurrent streaming 3D reconstruction, when processing input sequences containing thousands of frames, including whether reconstruction and pose estimation remain stable and accurate over kilometer-scale trajectories.

Background

The paper proposes MeMix, a training-free, plug-and-play state-update module designed to improve long-horizon stability for streaming 3D reconstruction by selectively updating only a subset of memory patches. Extensive experiments demonstrate gains on sequences up to several hundred frames across multiple datasets and backbones.

However, the authors explicitly note that MeMix has not been evaluated on extremely long sequences, such as thousands of frames corresponding to kilometer-scale trajectories, which are critical for applications like navigation and perception. This leaves the long-range behavior of MeMix an unresolved question.

References

Although MeMix surpasses previous main-stream methods in long-horizon inference, we have not tested what happens when the input contains thousands of frames.

— MeMix: Writing Less, Remembering More for Streaming 3D Reconstruction  (2603.15330 - Dong et al., 16 Mar 2026) in Conclusion, Limitations

Suppressing drift over kilometer-scale sequences remains an open challenge, addressed through test-time gradient updates, training-free memory management, long-range token pools, explicit spatial memory, stage-decoupled streaming, and offline global optimization, each trading off online capability against global consistency.

— Scal3R: Learning Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstruction  (2609.04201 - Lin et al., 3 Sep 2026) in Section 2, subsection “Long-sequence Streaming 3D Reconstruction”