🌐Project Page • 📄 Paper

This is the official repository for the ICCV 2025 paper:

GECO: Geometrically Consistent Embedding with Lightspeed Inference.

Regine Hartwig^1,2, Dominik Muhle^1,2, Riccardo Marin^1,2, Daniel Cremers^1,2,

¹Technical University of Munich, ²MCML

ICCV 2025 (arXiv)

If you find our work useful, please consider citing our paper:

@inproceedings{hartwig2025geco,
  title={GECO: Geometrically Consistent Embedding with Lightspeed Inference},
  author={Hartwig, Regine and Muhle, Dominik and Marin, Riccardo and Cremers, Daniel},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2025}
}

We address the task of geometry-aware feature encoding. A common way to test geometric awareness is through keypoint matching: Given a source image with an annotated keypoint, the goal is to predict the keypoint in the target image by selecting the location with the highest feature similarity. Below are two examples:

We introduce a training paradigm and a lightweight architecture for learning from image pairs with sparsely annotated keypoints. Additionally, we enhance the testing of features by introducing subdivisions of the commonly used PCK metric and a centroid clustering approach for more densely evaluating the feature space.

If you're using a Linux machine, set up the Python environment with:

conda create --name geco python=3.10
conda activate geco
bash setup_env.sh

To use Segment Anything (SAM) for mask extraction:

pip install git+https://github.com/facebookresearch/segment-anything.git
wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth

Install odise in case you want to run the Geo baseline

git clone [email protected]:NVlabs/ODISE.git
cd ODISE
pip install -e .

Run the following scripts to prepare each dataset:

• APK:

  bash download_data/prepare_apk.sh
  wget https://github.com/Junyi42/GeoAware-SC/blob/master/prepare_ap10k.ipynb

  Then run the notebook from GeoAware-SC to preprocess the data.
• CUB:

  bash download_data/prepare_cub.sh

• PascalParts

  bash download_data/prepare_pascalparts.sh

• PFPascal

  bash download_data/prepare_pfpascal.sh

• SPair-71k:

  bash download_data/prepare_spair.sh

1. Define <your_precomputed_masks_path> in the dataset config files.
2. Define <path-model-seg> in store_masks.yaml pointing to the path, where sam_vit_h_4b8939.pthis stored.
3. Select the datasets to process in store_masks.yaml.
4. Run:

    python scripts/store_masks.py --config-name=store_masks.yaml 

Choose a path and define <your_precomputed_feats_path> in the dataset config files. Define which dataset you want to extract the features for in store_feats.yaml and run

python scripts/store_feats.py --config-name=store_feats.yaml 

Pretrained weights are available in pretrained_weights/geco.

We provide interactive jupyter notebooks for testing.

• 📚 Data Loading Demo

  Validate dataset preparation and path setup.

• 🎨 Segmentation Demo

  Visualize part segmentation using a simple linear classifier.

• 📍 Keypoint Transfer Demo

  Explore keypoint transfer and interactive attention maps.

Run full evaluation:

python scripts/run_evaluation_all.py --config-name=eval_all.yaml 

Evaluate inference time and memory usage:

python scripts/run_evaluation_time_mem.py --config-name=eval_time_mem.yaml 

Evaluate segmentation metrics:

python scripts/run_evaluation_seg.py --config-name=eval_seg.yaml 

Before training, comment out the following block in configs/featurizer/dinov2lora.yaml:

init:
  id: geco
  eval_last: True

Then run:

python scripts/train_pairs.py --config-name=train_pairs.yaml