Overview
- Field: Computer Vision
- Authors: Chin-Yang Lin, Yang-Che Sun, Cheng Sun
- Published: 2026-09-03
- arXiv: 2609.04201
- Lightweight learnable tokens (~1% of parameters) are injected into a fully frozen backbone via asymmetric attention
- Poses are queried relative to multiple past keyframes rather than a single fixed anchor
- An online pose graph optimization system with loop closure suppresses long-range drift
- Converges in 8 hours on a single GPU
- Reduces average ATE by over 60% on KITTI compared to online baselines
- Achieves state-of-the-art performance on Virtual KITTI, Sintel, TUM-Dynamic, ScanNet, and 7-Scenes
Summary
Online 3D reconstruction models perform poorly on long videos. The reason is that regressing poses relative to a fixed first-frame anchor forces the model to extrapolate beyond the training distribution, and small drifts continuously accumulate and amplify into severe geometric collapse.
However, the authors observed that per-frame depth remains stable throughout this failure process—the backbone network's local geometry stays intact, and only the global pose head collapses. Based on this decoupling observation, they propose Scal3R, which reformulates online reconstruction as multi-reference relative pose querying.
Method
Results
Abstract
Online 3D reconstruction models perform poorly on long videos because regressing poses relative to a fixed first-frame anchor forces extrapolation far beyond the training distribution. We introduce Scal3R, which reformulates online reconstruction as multi-reference relative pose querying using lightweight learnable tokens (~1% of parameters) injected into a frozen backbone via asymmetric attention.
---
*Auto-collected on 2026-09-05*