capture lock: Modular Visual Odometry Research Platform

This repository contains a research-oriented visual odometry (VO) platform, built on top of COLMAP's geometric core, and designed to support modular substitution of key VO components. Our goal is to enable fast prototyping and benchmarking of advanced algebraic solvers — such as [DRaM-LHM], which is included as a case study.

Originally developed at the Flatiron Institute, this system is meant to provide a flexible framework for testing alternative formulations of:

• Perspective-n-Point (currently demonstrated with via DRaM-LHM in this repo)
• Essential/Fundamental matrix estimation
• Bundle adjustment
• And more

---

Architecture

• workspace — Coordinates the VO pipeline and allows component-wise benchmarking.
• pnp/ — Includes point sampling, synthetic pair generation, and scripts for PnP evaluation.
• estimate/ — Contains all implemented geometric estimators. New methods can be added easily.
• calib_loader/ — Provides dataset-specific calibration loading. Supports plug-and-play dataset handling.

This layout makes it easy to replace individual modules (e.g., PnP solver or feature backend) for targeted testing or integration into larger SLAM systems.

🚀 Getting Started

This system is designed to run inside a Docker container for reproducibility. We use the prebuilt image cyidocker/colmap:v02.

1. Launch the Docker container

   Mount your local codebase and datasets folder:

   docker run -it \
     -v /your/path/to/capture_lock/:/tmp1 \
     -v /your/path/to/datasets/:/tmp3 \
     --ipc=host --shm-size=16g \
     -p 8096:8096 \
     --rm cyidocker/colmap:v02 /bin/bash

2. Build the project inside the container

      cd /tmp1/build
      make -j32

3. Run PnP Experiments

      cd /workspace/pnp
      ./pnp [data_generator_option] [method_option] [ransac_option]

Related Publications and Acknowledgements

The iterative PnP solver, DRaM-LHM:

DRaM-LHM: A Quaternion Framework for Iterative Camera Pose Estimation
Chen Lin, Weizhi Du, Zhixiang Min, Baochen She, Enrique Dunn, Sonya M. Hanson
[ICCV 2025]

This project builds on the foundational theory of adjugate quaternions:

Exploring the adjugate matrix approach to quaternion pose extraction
Andrew Hanson and Sonya Hanson
Arxiv (https://arxiv.org/abs/2205.09116)

We gratefully acknowledge the open-source contributions used in our system:

• UPnP and DLS: Implemented via [OpenCV]
• EPnP: From [COLMAP]
• SQPnP: [Original author’s implementation]
• LHM: Rewritten from a [C version] into customized C++ code
• REPPnP: [Open-source C++ implementation]

We sincerely thank the authors and maintainers of these implementations for their contributions to the research community.