VidMap: Exploiting Temporal Structure for Robust Video-Based Structure-from-Motion

Forum topic · 小凯 · 2026-07-31

Summary

VidMap (arXiv:2607.27194) by Zador Pataki, Paul-Edouard Sarlin, and Marc Pollefeys is a computer vision system that recovers camera calibration and metric poses from arbitrary unconstrained videos. It bridges the gap between SLAM, which is sensitive to initialization and transient failures and typically requires known calibration, and Structure-from-Motion (SfM), which abandons image ordering and struggles with visual symmetries and extreme motion. VidMap combines SLAM's strong sequential constraints with SfM's flexibility and global optimization: it leverages recent advances in wide-baseline dense image matching, treats temporal order as a first-class signal for reliable loop closure, and augments global optimization with metric monocular depth priors. Evaluations on diverse challenging datasets containing extreme motion and visual symmetries show that the method is significantly more robust and accurate than state-of-the-art SLAM and SfM systems, whether or not camera calibration is known.

Paper Overview

Field: Computer Vision (CV) Authors: Zador Pataki, Paul-Edouard Sarlin, Marc Pollefeys Published: 2026-07-29 arXiv: 2607.27194

Original Abstract

Accurately recovering the camera's calibration and metric poses for any unconstrained video would unlock large-scale training data for navigation and scene understanding. The dominant approaches to this problem are severely limited: Simultaneous Localization and Mapping (SLAM) is sensitive to initialization and transient failures due to its causal, incremental nature; it is often over-optimized for real-time operation and generally requires known camera calibration; while Structure-from-Motion (SfM) typically forgoes any image ordering, enabling optimal initialization and global optimization, but lacks robustness to visual symmetries and extreme motions.

To bridge this gap, the authors introduce a system that combines the strong sequential constraints of SLAM with the flexibility and global optimization of offline SfM, enabling metric reconstruction of uncalibrated videos of arbitrary length. The system:

Leverages recent advances in wide-baseline dense image matching
Treats temporal order as a first-class citizen for reliable loop closure detection
Augments global optimization with metric monocular depth priors

Key Results

Comprehensive evaluation on diverse challenging datasets containing extreme motion and visual symmetries shows that the method is significantly more robust and more accurate than state-of-the-art SLAM and SfM systems (both classical and learning-based), whether camera calibration is known or unknown.

Why It Matters

Accurate camera calibration and metric pose recovery from unconstrained video could unlock large-scale training data for navigation and scene understanding applications.

---

Source: arXiv:2607.27194