cvg/vidmap

VidMap: Exploiting Temporal Structure for Video-Based Structure-from-Motion (ECCV 2026)

What it solves

VidMap is an offline Structure-from-Motion (SfM) system designed to reconstruct 3D environments from video sequences. It addresses the challenge of accurately estimating camera poses, camera intrinsics, and sparse 3D maps by leveraging the inherent temporal continuity found in video data.

How it works

The system processes video through a two-stage pipeline: a video-aware frontend and a provenance-aware global mapper. The frontend extracts temporal tracks, performs keyframe selection, and estimates depth, while the global mapper uses these inputs to perform global optimization and bundle adjustment. It builds upon established tools like COLMAP and supports GPU acceleration for the optimization stages to improve performance.

Who it’s for

This project is for researchers and computer vision practitioners working on 3D reconstruction, SLAM (Simultaneous Localization and Mapping), and visual odometry who need a robust, offline system to process video datasets.

Highlights

  • Leverages temporal structure to improve reconstruction accuracy over standard image-based SfM.
  • Includes a flexible configuration system using Hydra for fine-grained control over the pipeline.
  • Supports GPU-accelerated global positioning and bundle adjustment for faster processing.
  • Provides built-in visualization tools, including browser-based viewers and Rerun integration for playback of solver states and 3D flythroughs.
  • Includes comprehensive benchmarking support for standard datasets like LaMAR, EuRoC, and ETH3D-SLAM.

Related

  • Project
    mapillary/OpenSfM

    A Structure from Motion library in Python that reconstructs 3D scenes and camera poses from multiple images, integrating sensor data for geographical alignment.

  • Project
    colmap/colmap

    COLMAP is an open‑source Structure‑from‑Motion and Multi‑View Stereo pipeline that automatically turns collections of photos into 3‑D models. It offers a GUI, command‑line tools, Python bindings, GPU‑accelerated feature extraction, and cross‑platform binaries, making it a go‑to foundation for research and applications in computer vision, robotics, AR/VR, and any workflow that needs camera poses and dense scene geometry.

  • Project
    borglab/gtsfm

    GTSfM is a modular, parallelized end-to-end Structure-from-Motion pipeline that reconstructs 3D scenes from images using distributed computing.

  • Project
    cvg/limap

    A toolbox for holistic 3D mapping and structure from motion that jointly optimizes points, lines, planes, and parametric primitives for more accurate 3D reconstructions.

  • Project
    cre185/InstantSfM

    A GPU-native Structure from Motion (SfM) pipeline that accelerates the process of reconstructing 3D structures from 2D images, including integrated support for 3D Gaussian Splatting.