LoGeR scales feed-forward dense reconstruction to very long videos without any optimisation afterwards. It processes the stream in chunks with bidirectional attention inside each chunk, and links chunks with a learned hybrid memory: a test-time-training memory that anchors the global frame and stops scale drift, plus sliding-window attention that keeps uncompressed context for aligning neighbouring chunks.
Findings
- Trained on 128-frame sequences, it generalises to thousands of frames at inference (abstract).
- Over 74% lower ATE on KITTI than prior feed-forward methods; tested on a repurposed VBR dataset with sequences of up to 19k frames.
Limits
- Abstract only so far. In-air driving and handheld data.
Relevance
Robot videos run to many thousands of frames, so onboard use depends on chunking with drift control. Whether a learned memory notices revisiting a place, the way a loop closure would, is the open question for repeated runs. See foundation models for robotics and wang-2025-vggt.
Links
- arXiv · code · project page
- Subfield: Mapping and localisation