This repository presents our solution to the Image Matching Challenge 2022, a competition hosted by Google Research as part of the CVPR'22 Workshop: Image Matching: Local Features & Beyond. Our solution finished at 34th place out of 642 participating teams, earning us a Kaggle Competition Silver Medal 🥈.
We hope our solution can serve as a resource for the community and inspire future work. If you find this repository useful, please consider giving it a star!
Authors: Armando Fortes & David Pissarra
Our solution leverages the SuperGlue (detector-based) and LoFTR (detector-free), two highly regarded methods for the image matching task. The chosen feature detector to complement the SuperGlue pre-trained network was SuperPoint. In addition, we add QuadTree Attention to LoFTR, improving the performance of the detector-free model.
For each pair of input images, we feed them into the two different local feature matching heads (i.e. the detector-based and detector-free approaches described above). The resulting matches from both heads are then concatenated and fed into the MAGSAC consensus algorithm for inlier/outlier detection and fundamental matrix calculation.
More details on the implementation and experiments can be found in our report.
Some inference examples, using images from famous monuments around the world. The colors of the resulting matches range from green for high-confidence matches to red for low-confidence matches.
| Colosseum | Taj Mahal |
 |
 |
| Basilica of the Sagrada Familia | Trevi Fountain |
 |
 |