XDG: Accelerated Visual Disambiguation (WACV 2027)

Gonglin Chen, Ben Southall, Hanyuan Xiao, Wenbin Teng, Haolin Xiong, Tianwen Fu, Junyi Ouyang, Kshitij Singh Minhas, Supun Samarasekera, Rakesh Kumar, Yajie Zhao

Project Page

Table of Contents

• Overview
• Updates
• Installation
• Checkpoints
• Inference
  • Filter a COLMAP database
  • Score an explicit pair list
• Training
• Evaluation
• Citation
• License
• Acknowledgements
• Third-party code

Overview

XDG detects and removes visually plausible but geometrically inconsistent image pairs before SfM. By filtering these false-match edges, it prevents corrupted reconstructions and recovers more accurate camera poses. Across pairwise and SfM benchmarks, XDG provides comparable disambiguation performance to DG++ while running 3.5× faster.

Updates

[08/30/2026] Code and pretrained checkpoint released.

[08/10/2026] XDG has been accepted to WACV 2027.

Installation

Clone the repository, then create and activate the provided Conda environment:

git clone https://github.com/xtcpete/xdg.git
cd xdg
conda env create -f environment.yml
conda activate xdg

Alternatively, use a Python virtual environment:

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt
python -m pip install -e Depth-Anything-3

You can also build a docker image with:

docker build -t xdg .

Mount image, weight, and output directories when running the container; these directories are intentionally excluded from the image build context.

Checkpoints

Testing and inference automatically load the released checkpoint from Hugging Face. The checkpoint is also available on Google Drive.

Inference

Inference uses only the model configuration at configs/model_configs/xdg.yaml. The file contains the released architecture and the preprocessing/runtime defaults used at inference time.

Example hloc reconstruction

pipeline.py runs ALIKED + LightGlue matching, geometric verification, XDG filtering, and mapping. It uses exhaustive pairs for fewer than 30 images and NetVLAD retrieval otherwise. Install hloc using python -m pip install git+https://github.com/cvg/Hierarchical-Localization.git if it is not already installed.

python pipeline.py \
  --images path/to/images \
  --outputs outputs/example \
  --threshold 0.8

Use an empty output directory for each run. The reconstruction is saved under outputs/example/sfm/, alongside the original and filtered COLMAP databases. Pair probabilities are saved under outputs/example/filtering/. Use --feature_conf and --matcher_conf to select compatible hloc configurations; python pipeline.py --help lists the other options.

For better feature matching, check out our previous work RDD and its SfM example.

Filter a COLMAP database

The database must already contain images and geometrically verified pairs in two_view_geometries.

python remove_doppelgangers.py \
  --config configs/model_configs/xdg.yaml \
  --database_path path/to/database.db \
  --input_image_path path/to/images \
  --output_path output/scene \
  --threshold 0.8

This writes pair probabilities to output/scene/pair_probability_list.npy and creates a filtered copy such as database_threshold_0.800.db. The input database is not modified. Then run COLMAP mapper with filtered database.

Score an explicit pair list

Each non-empty line in the pair file must contain two whitespace-separated image paths relative to --input_image_path:

image_0001.jpg image_0002.jpg
subdir/image_0003.jpg subdir/image_0004.jpg

Run inference without a COLMAP database:

python remove_doppelgangers.py \
  --config configs/model_configs/xdg.yaml \
  --pairs_txt path/to/pairs.txt \
  --input_image_path path/to/images \
  --output_path output/pairs

--batch_size, --img_size, --mode, --num_workers, and --device can override the model-config defaults.

Training

Training uses configs/training_configs/training.yaml. Its model_config field points to the standalone model configuration, so architecture settings have a single source of truth. At training startup, model.backbone.model_path initializes the DA3 backbone from pretrained weights.

The dataset configuration expects NumPy pair metadata files. Each row starts with the two image paths and a binary label; additional metadata columns are allowed. Please follow doppelgangers++ for downloading the datasets. Update the paths under data.train and data.test for your local datasets, then run:

python train.py configs/training_configs/training.yaml 

Resume training with --resume_from path/to/last.ckpt. Override the configured epoch count with --max_epochs.

Evaluation

Validation requires the training config because it needs dataset settings as well as the referenced model config:

python test.py

The released checkpoint is loaded automatically. Use --ckpt path/to/xdg.pth for a different local checkpoint.

The Doppelgangers test set is used by default. Evaluate on VisymScenes with --dataset visymscenes; its paths are configured under data.test_sets in the training YAML.

The command reports average precision, ROC AUC, operating-point precision/recall, and inference time per pair.

Citation

If you find XDG useful in your research, please cite:

@misc{chen2026xdgacceleratedvisualdisambiguation,
  title={XDG: Accelerated Visual Disambiguation},
  author={Gonglin Chen and Ben Southall and Hanyuan Xiao and Wenbin Teng and Haolin Xiong and Tianwen Fu and Junyi Ouyang and Kshitij Singh Minhas and Supun Samarasekera and Rakesh Kumar and Yajie Zhao},
  year={2026},
  eprint={2608.29733},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2608.29733},
}

License

This project is licensed under the Apache License 2.0. The license applies to the original code in this repository and checkpoint. Datasets referenced for training and evaluation are subject to their respective providers' licenses and terms;

Acknowledgements

We thank the authors of these great repositories: Depth Anything 3, Doppelgangers, Doppelgangers++, and COLMAP, along with many other inspiring works from the community.

This material is based upon work supported by the Intelligence Advanced Research Projects Activity under prime Contract No. 140D0423C0034. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government.

Third-party code

Depth Anything 3 is vendored under Depth-Anything-3/ and retains its own license in Depth-Anything-3/LICENSE.