👉 Kaggle_Image_Matching_Challenge_24

This repository contains our solution for the "Image Matching Challenge 2024 - Hexathlon" Kaggle competition, where we achieved a 🥈silver medal.

🤠 Algorithm Overview

🧐 Algorithm Details

🧐 Our approach consists of three main stages:

1. Image Retrieval

We retrieve images from various 3D scene datasets using pre-trained EfficientNet-B6 & B7 models from ImageNet to extract image features. The cosine distance metric is used to rank images based on similarity, selecting the top n images for each scene.

2. Feature Extraction and Matching

For the retrieved images, we employ two parallel pipelines to extract and match keypoints:

• Corner Detection with AdaLAM: We use Kornia's CornerGFTT feature detector to extract keypoints and apply the AdaLAM algorithm for robust feature matching. Successful match pairs are stored.
• SuperPoint & SuperGlue: We extract keypoints using SuperPoint and perform feature matching with SuperGlue, retaining the valid match pairs.

3. 3D Pose Estimation

We merge the successful match pairs from both pipelines, remove duplicates, and use PyCOLMAP to compute the final 3D spatial relationships, including camera positions and pose estimation.

Model References

• SuperPoint & SuperGlue (source code and paper)
• GFTT Algorithm Explanation
• Kornia Documentation

Citation

If you find our work helpful, please consider citing:

@inproceedings{sarlin20superglue,
  author    = {Paul-Edouard Sarlin and
               Daniel DeTone and
               Tomasz Malisiewicz and
               Andrew Rabinovich},
  title     = {{SuperGlue}: Learning Feature Matching with Graph Neural Networks},
  booktitle = {CVPR},
  year      = {2020},
  url       = {https://arxiv.org/abs/1911.11763}
}

Appreciate your interest in our work!