SceneBound: How Much Should a Streaming 3D Reconstruction Model Remember?
SceneBound: How Much Should a Streaming 3D Reconstruction Model Remember?
This paper introduces SceneBound, a training-free rule for streaming 3D reconstruction that determines when to write frames based on camera pose distance from stored keyframes. The method keeps the keyframe count within a constant factor of the covering number of visited poses and achieves the lowest mean trajectory error among nine evaluated training-free rules across four models and three memory types, with keyframe growth between 3% and 6%.
The authors propose SceneBound, a training-free rule for streaming 3D reconstruction that writes frames only when the predicted camera pose exceeds a fixed radius from stored keyframes, achieving lower mean trajectory error than nine other published rules across four models and three memory types.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
memory types
3
View original evidence
poses, the fewest views that keep a keyframe near every camera, so the count grows with the places visited and stays flat when the camera returns. In large scenes, SceneBound closes a submap once decoding its first frame again reveals drift. SceneBound uses one radius across four released models and three memory types. In the recurrent state of CUT3R and the KV cache of StreamVGGT, it achieves the lowest mean trajec
Open source S4gates compared
9
View original evidence
ng the update budget of a gate at uniformly spaced frames, with no signal, matches nine published gates, and the budget bounds the drift of the state wherever the writes fall. We propose SceneBound, a training-free rule that writes a frame only when its predicted camera pose lies at least a fixed radius from every stored keyframe. Its keyframe count stays within a constant factor of the covering number of the visited
Open source S3scene passes evaluated
1
Reported trials
View original evidence
tory error among the published training-free rules we evaluate, in one pass and over nine passes of a scene, where its keyframes grow by 3% to 6%.
Open source S5scene passes evaluated
>9
Reported trials
View original evidence
tory error among the published training-free rules we evaluate, in one pass and over nine passes of a scene, where its keyframes grow by 3% to 6%.
Open source S5
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
tory error among the published training-free rules we evaluate, in one pass and over nine passes of a scene, where its keyframes grow by 3% to 6%.
Open source S5
poses, the fewest views that keep a keyframe near every camera, so the count grows with the places visited and stays flat when the camera returns. In large scenes, SceneBound closes a submap once decoding its first frame again reveals drift. SceneBound uses one radius across four released models and three memory types. In the recurrent state of CUT3R and the KV cache of StreamVGGT, it achieves the lowest mean trajec
Open source S4
ng the update budget of a gate at uniformly spaced frames, with no signal, matches nine published gates, and the budget bounds the drift of the state wherever the writes fall. We propose SceneBound, a training-free rule that writes a frame only when its predicted camera pose lies at least a fixed radius from every stored keyframe. Its keyframe count stays within a constant factor of the covering number of the visited
Open source S3
Source：arXiv Robotics — research abstracts · arxiv.org