nianticlabs/map-free-reloc

[ECCV 2022] Map-free Visual Relocalization: Metric Pose Relative to a Single Image

What it solves

This project addresses the challenge of visual relocalization—determining the precise location and orientation of a camera in a 3D space. Traditional methods typically require building a complex 3D map of a scene using hundreds of images. This project enables "map-free" relocalization, allowing a system to determine its metric pose relative to a single reference image without needing a pre-built 3D map.

How it works

The project provides a reference implementation for the research presented at ECCV 2022. It includes a large-scale dataset of 655 outdoor scenes, each featuring a specific point of interest. The system uses query images and reference images to estimate the camera's pose. It supports various baseline approaches, including Relative Pose Regression (RPR) and feature matching combined with depth estimation, and provides tools for evaluating and visualizing these pose estimates against ground-truth data.

Who it’s for

This repository is designed for computer vision researchers and developers working on spatial AI, robotics, and augmented reality who need to benchmark algorithms for camera pose estimation in scenarios where 3D map data is unavailable.

Highlights

  • Includes a dedicated dataset of 655 outdoor scenes for training and testing.
  • Provides an official online benchmark and leaderboard for comparing different relocalization methods.
  • Supports both single-frame and multi-frame (up to 9 frames) query tracks.
  • Offers modular configuration for testing different encoders, aggregators, and loss functions.
  • Includes pre-trained models and scripts for local evaluation and visualization of results.

Related

  • Project
    zju3dv/OnePose_Plus_Plus

    OnePose++ is a keypoint-free, one-shot object pose estimation system that determines an object's 3D position and orientation without requiring CAD models.

  • Project
    facebookresearch/mvdust3r

    A single-stage 3D scene reconstruction tool that creates 3D point clouds and camera poses from sparse RGB views in about 2 seconds without requiring pre-known camera poses.

  • Project
    PoseLib/PoseLib

    A library of minimal solvers and robust estimators for camera pose estimation, supporting various correspondences and camera models for fast 3D reconstruction.

  • Project
    apple-aiml-research/ml-matrix3d

    Matrix3D is a unified photogrammetry model that performs pose estimation, depth prediction, and novel view synthesis to enable 3D reconstruction from single or few-shot unposed images.

  • Project
    perfanalytics/pose2sim

    An open-source workflow for markerless 3D kinematics that converts 2D pose estimation from standard cameras into research-grade 3D joint angles and skeletal analysis.