Computer Vision / Competition Project

Kaggle Silver Medal: Image Matching Challenge 2022

Built on top of a LoFTR image matching baseline, then improved inference flow and geometric verification parameters to push the Kaggle Image Matching Challenge 2022 score to 0.726 and earn a silver medal.

Competition

Image Matching 2022

Placement

40 / 642 teams

Best Score

0.726

Core Stack

LoFTR + MAGSAC

Problem Statement

The competition required robust image matching and camera-geometry reasoning across large viewpoint changes, lighting variation, occlusion, and repeated structures.

Key Finding

The performance gain did not come from swapping in a new model, but from stabilizing and tuning the full LoFTR pipeline: image resizing, grayscale input, MAGSAC parameters, and submission logic.

Why This Project Matters

This project carries strong external proof. A Kaggle silver medal and a 40 / 642 finish show not just model usage, but the ability to tune, validate, and improve under real leaderboard pressure.

For the portfolio, it works best as a high-technical-depth case study that proves comfort with complex computer vision workflows.

Technical Approach

The notebook centers on Kornia's pretrained LoFTR for correspondence generation, followed by OpenCV USAC_MAGSAC to estimate the fundamental matrix.

You also ran a compact parameter study across threshold, confidence, and iteration settings, ultimately pushing the score to 0.726.

How To Present It Honestly And Strongly

The strongest framing is not to claim a from-scratch invention, but to show that you took a strong LoFTR baseline, understood it quickly, and improved it effectively with real results.

That framing actually feels more like real work and better demonstrates the ability to understand existing systems, refine a pipeline, and move performance meaningfully.

Case Study

A fuller project breakdown for portfolio and interview storytelling.

Dataset

Competition-provided multi-view image pairs from the Kaggle Image Matching Challenge 2022 test set.
The task centered on predicting a valid fundamental matrix for each sample under challenging conditions such as viewpoint change, repeated structure, and visually ambiguous scenes.
The implemented notebook operated on Kaggle-hosted competition files and processed pairs one sample at a time for submission generation.

Methodology

Use a pretrained outdoor LoFTR matcher from Kornia to generate dense correspondences between image pairs.
Resize each image so the maximum side becomes 840 pixels, convert to grayscale, and run inference on GPU for the matching stage.
Estimate the fundamental matrix using OpenCV USAC_MAGSAC, then tune threshold, confidence, and iteration settings through a small parametric study.

Results

The final Kaggle notebook reached a score of 0.726.
The team placed 40th out of 642 teams and earned a silver medal.
The clearest gain documented in the notebook came from MAGSAC parameter tuning, where the best tested setting improved the score beyond the baseline 0.723 configuration.

Validation and Limitations

This notebook was built on top of an existing LoFTR baseline rather than from a blank slate, so the portfolio framing should emphasize optimization, adaptation, and experimentation.
The current notebook is competition-focused and not yet packaged like a clean production repository.
The existing version documents inference and tuning, but it would benefit from additional visual examples and a clearer ablation-style summary.

Recommendations

Present the project as a strong example of taking an advanced baseline, understanding its moving parts, and improving leaderboard performance through systematic tuning.
Add one pipeline diagram and a short explanation of where LoFTR ends and geometric verification begins, so the method is easier to understand for recruiters.
Include qualitative match visualizations to show how inlier filtering improves correspondence quality in difficult scenes.

Future Directions

Repackage the original Kaggle notebook into a cleaner repository with organized scripts, assets, and reproducible settings.
Add a post-competition visual analysis comparing raw LoFTR matches versus inliers retained after MAGSAC filtering.
Extend the project with a short comparison against another matching approach to show broader experimentation beyond one baseline.