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 specifically for video. It addresses the challenge of estimating camera poses, camera intrinsics, and creating a sparse 3D map from video sequences by leveraging the temporal structure inherent in video data.

How it works

VidMap uses a two-stage pipeline consisting of a video-aware frontend and a provenance-aware global mapping stage. It combines several key components:

  • Temporal Tracks: It tracks features across frames to maintain consistency.
  • Loop Closures: It identifies when the camera returns to a previously visited location to correct drift.
  • Metric Depth: It incorporates depth predictions to provide scale and geometry.
  • Global Optimization: It extends the COLMAP global mapping pipeline to optimize camera poses and 3D points.

Who it’s for

This tool is for researchers and developers working in computer vision, 3D reconstruction, and robotics, specifically those needing to convert video footage into accurate 3D maps and camera trajectories.

Highlights

  • COLMAP Integration: Extends the existing COLMAP global mapping pipeline for improved video-based SfM.
  • GPU Acceleration: Supports CUDA and cuDSS via Ceres 2.3+ for faster global positioning and bundle adjustment.
  • Comprehensive Visualization: Includes a browser-based viewer and Rerun-based playback for solver states and flythroughs.
  • Benchmark Support: Built-in tools to prepare and evaluate results on datasets like LaMAR, EuRoC, ETH3D-SLAM, and CroCoDL.
  • Flexible Configuration: Uses Hydra and OmegaConf for detailed control over frontend and mapping hyperparameters.

Related

  • Project
    colmap/colmap

    A general-purpose Structure-from-Motion and Multi-View Stereo pipeline for reconstructing 3D structures from image collections.

  • 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
    borglab/gtsfm

    An end-to-end, distributed Structure-from-Motion pipeline based on GTSAM and Dask that reconstructs 3D scenes from 2D images with native parallel computation support.

  • 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.

  • Project
    ACFR-RPG/DynOSAM

    DynoSAM is a Stereo/RGB-D Visual Odometry pipeline for Dynamic SLAM that estimates camera poses and object motions while mapping static backgrounds.