Official Code for ArgMatch

November 23, 2025 · View on GitHub

This repository contains the official implementation of the ICCV 2025 paper:

ArgMatch: Adaptive Refinement Gathering for Efficient Dense Matching by Yuxin Deng, Kaining Zhang, Linfeng Tang, Jiaqi Yang, and Jiayi Ma.

Installation

1. Clone the repository

git clone https://github.com/ACuOoOoO/ArgMatch.git
cd ArgMatch

2. Create the Conda environment

Use the provided env.yaml file to create the Conda environment:

conda env create -f env.yaml

Then activate the environment:

conda activate argmatch

3. Install NATTEN

After activating the environment, install the NATTEN library:

pip3 install natten==0.17.1+torch230cu121 -f https://shi-labs.com/natten/wheels

Warning: ArgMatch relies on NATTEN’s unfused neighborhood attention API (the deprecated interface for computing query–key similarities). Recent versions of natten have removed this API, so you must use the compatible combination torch==2.3.0 and natten==0.17.1+torch230cu121. We are working on a new version of ArgMatch that removes this dependency and will keep the codebase updated with future NATTEN releases.

Pretrained Checkpoint

Download the pretrained ArgMatch checkpoint from:

After downloading, place the checkpoint file ckpt_clean.pth in:

ArgMatch/workspace/checkpoints/

The final path should look like:

ArgMatch/
├── workspace/
│   ├── checkpoints/
│   │   ├── ckpt_clean.pth

Data Preparation

Download MegaDepth and ScanNet test following the instructions from:

This should result in the following data structure:

ArgMatch/
├── data/
│   ├── megadepth/
│   │   ├── phoenix/
│   │   ├── Undistorted_SfM/
│   │   ├── prep_scene_info/
│   │   ├── 0015_0.1_0.3.npz ...
│   ├── scannet/
│   │   ├── scans_test/
│   │   ├── scannet_eval_list.txt
│   │   ├── test.npz

Test

Demo

We provide a simple demo script experiments/demo.py that:

  • loads the pretrained ArgMatch checkpoint,
  • takes two input images,
  • computes dense correspondences,
  • warps image B to image A, and
  • saves visualizations (flow, certainty map, and concatenated images).

Run the demo with:

python -m experiments.demo \
  --checkpoint workspace/checkpoints/ckpt_clean.pth \
  --image_A assets/im_A.jpg \
  --image_B assets/im_B.jpg \
  --output_dir assets/outputs

After running, you should find the following files in assets/outputs/:

  • flow.png – visualization of the dense flow field
  • certainty.png – certainty / confidence map
  • concat.png – concatenation of image A, image B, and the warped image B

Evaluation

Before running the evaluation scripts, please create a directory for storing the results:

mkdir -p results

To reproduce the relative camera pose estimation results on MegaDepth-1500:

python -m experiments.test \
  --gpu 0 \
  --tests test_mega1500 \
  --sample_thresh 0.03

To reproduce the relative camera pose estimation results on ScanNet:

python -m experiments.test \
  --gpu 1 \
  --tests test_scannet \
  --sample_thresh 0.1

To reproduce the dense matching results on MegaDepth-1500:

python -m experiments.test \
  --gpu 2 \
  --tests test_megdepth_dense_scale

The corresponding logs and metrics will be saved under the results/ directory.

Training

We will release the training code in a future update.

Citation

If you find this repository useful in your research, please consider citing:

@inproceedings{deng2025argmatch,
  title     = {ArgMatch: Adaptive Refinement Gathering for Efficient Dense Matching},
  author    = {Yuxin Deng and Kaining Zhang and Linfeng Tang and Jiaqi Yang and Jiayi Ma},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2025}
}

Acknowledgements

This implementation is heavily inspired by and builds upon the excellent work of DKM and ROMA. We sincerely thank the authors and maintainers of these projects for open-sourcing their code and ideas.

License

This project is released under the terms of the license specified in the LICENSE file.