GitHub

Jinjiang You*, Hewei Wang*, Yijie Li, Mingxiao Huo, Long Vân Tran Ha, Mingyuan Ma, Jinfeng Xu, Jiayi Zhang, Puzhen Wu, Shubham Garg, Wei Pu

[Paper (IEEE Xplore)] [arXiv] [BibTeX] [Gallery]

Multi-Cali Anything is a dense-feature-driven multi-frame camera calibration method designed for large-scale camera arrays. Unlike traditional calibration methods that require dedicated checkerboard captures, our approach directly refines camera intrinsics from scene data, eliminating the necessity for additional calibration captures.

🔍 Why Use This?

  • ✅ No extra calibration captures needed – Uses scene data instead of calibration patterns.
  • ✅ Seamless integration – Works as an add-on to existing SfM pipelines (e.g., COLMAP, Pixel-Perfect SfM).
  • ✅ Dense feature refinement – Reduces keypoint errors for higher calibration accuracy.
  • ✅ Multi-frame optimization – Ensures consistent intrinsics across multiple frames.
  • ✅ Efficient processing - Suitable for large-scale camera arrays with multiple frame captures.
  • ✅ High accuracy - Achieves nearly the same precision as dedicated calibration processes.

🛠 How It Works?

  • 1️⃣ Run any SfM pipeline (COLMAP, Pixel-Perfect SfM, etc.) to obtain initial sparse reconstructions (camera parameters + sparse 3D models).
  • 2️⃣ Use Multi-Cali Anything to get refined camera intrinsics and 3D models.

Installation

We recommend running the project in a docker container. All dependencies will be installed if you build the docker image using our Dockerfile.

Git clone the repository to the local machine:

cd /path/to/your/workspace
git clone https://github.com/YJJfish/Multi-Cali-Anything.git
cd Multi-Cali-Anything

Choose the right CUDA_ARCHITECTURES for your GPU according to https://developer.nvidia.com/cuda-gpus, and set the value in Dockerfile:

# File: Multi-Cali-Anything/Dockerfile
# TODO: Choose the right CUDA_ARCHITECTURES for your GPU according to https://developer.nvidia.com/cuda-gpus.
# Some versions may not be supported by CUDA 11.7.1. E.g. 89.
ARG CUDA_ARCHITECTURES=86

Also set the value in src/CMakeLists.txt:

# File: Multi-Cali-Anything/src/CMakeLists.txt
set_target_properties(calibration PROPERTIES CUDA_ARCHITECTURES "86")

Build a docker image with name calibration and tag 1.0, using the provided Dockerfile:

docker build -t calibration:1.0 .

Run a Docker container with the following options:

  • Mount the workspace folder in the host machine to /home/ubuntu/workspace in the container.
  • Specify the amount of RAM available to the container (e.g., 16GB) according to your dataset.
  • Specify the amount of shared memory available to the container (e.g., 8GB) according to your dataset.
  • Include GPU support for the container by using --gpus all.
cd /path/to/your/workspace
docker run -dit \
	--name calibration-container \
	-v $(pwd):/home/ubuntu/workspace \
	--memory=16g \
	--shm-size=8g \
	--gpus all \
	calibration:1.0
docker exec -it calibration-container bash

Inside the container, use CMake to generate the project files and use make to compile the project:

cd /home/ubuntu/workspace/Multi-Cali-Anything/src
mkdir build
cd build
cmake -S .. -B . -DCMAKE_BUILD_TYPE=Release
make

Tutorial

Make sure you are running these commands in the docker container environment.

Dataset

Download the Multiface dataset (or its mini-dataset).

cd /home/ubuntu/workspace
git clone https://github.com/facebookresearch/multiface
cd multiface
python3 download_dataset.py --dest "../mini_dataset/" --download_config "./mini_download_config.json"

Data Preprocessing

Go back to the Multi-Cali Anything repo directory.

cd /home/ubuntu/workspace/Multi-Cali-Anything

The images of Multiface dataset are gathered per camera. However, for SfM applications, images are required to be gathered per frame.

Reorder the dataset images using script/gather_images.py:

mkdir ../E057
python3 script/gather_images.py \
	../mini_dataset/m--20180227--0000--6795937--GHS/images/E057_Cheeks_Puffed \
	../E057/dataset

Also copy the KRT (which contains the ground-truth extrinsics) file to the same folder as the reordered dataset.

cp ../mini_dataset/m--20180227--0000--6795937--GHS/KRT ../E057/KRT

Our project functions as an add-on to other SfM pipelines. It uses the sparse reconstruction results of other SfM pipelines and refines the models. We provide colmap_batch.py or pixelsfm_batch.py to run COLMAP or Pixel-Perfect SfM. We use pixelsfm_batch.py as an example:

python3 script/pixelsfm_batch.py \
	../E057/dataset \
	../E057/pixelsfm

Optionally, run extract_dense_features.py to extract dense features and store them in database. These features will be used to compute featuremetric costs in our objective function if you enable dense feature refinement.

python3 script/extract_dense_features.py \
	../E057/dataset \
	../E057/pixelsfm \
	../E057/dense_features.sqlite \
	--mode pixelsfm

Finally, run our project:

./src/build/calibration \
	../E057/pixelsfm \
	../E057/dense_features.sqlite \
	../E057/KRT \
	--mode pixelsfm \
	--output ../E057/output

Citing

If you find our work useful in your research, please consider citing our paper.

@inproceedings{you2025multicalianything,
  author={You, Jinjiang and Wang, Hewei and Li, Yijie and Huo, Mingxiao and Ha, Long Vân Tran and Ma, Mingyuan and Xu, Jinfeng and Zhang, Jiayi and Wu, Puzhen and Garg, Shubham and Pu, Wei},
  booktitle={2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  title={Multi-Cali Anything: Dense Feature Multi-Frame Structure-from-Motion for Large-Scale Camera Array Calibration},
  year={2025},
  volume={},
  number={},
  pages={16106--16113},
  keywords={Three-dimensional displays;Temperature;Pipelines;Robot vision systems;Cameras;Distortion;Calibration;Optimization;Intelligent robots;Lenses},
  doi={10.1109/IROS60139.2025.11246221}}

Gallery

Visualization of reprojection errors, using intrinsics produced by different methods.

Reconstructions by multi-view stereo, using intrinsics produced by different methods. The reconstructions are compared against the ground-truth models, with blue indicating positive distances, red indicating negative distances, green indicating near-zerodeviation.

Reconstructions by DUSt3R, using intrinsics produced by different methods.

Read the original on github.com ↗