UPDATE - JOIN THE DISCORD:
https://discord.gg/yf4Ky2VXg




Modern cameras and lenses introduce predictable barrel distortion.

In professional, human-based photo editing workflows (Adobe Lightroom), this is typically fixed using lens profiles released and maintained by manufacturers or curated databases.

In practice, open-source lens profile coverage lags behind new camera/lens releases.

The problem: uncorrected distortion shows up as bent lines, warped interiors, and "off" perspective - especially noticeable in real-estate photography where straight edges matter.

If we lack a lens profile in our database, the customer get's back photos with barrel distortion.

The solution: one model that can apply lens correction to any photo - without needing the lens profile from the manufacturer!

**Your task in this competition is to automatically correct lens distortion in photos when no reliable lens profile is available.**

You'll receive a set of distorted test images. Using the provided training pairs (distorted originals + their corrected counterparts), build a model or algorithm that produces lens-corrected output images. Your corrected images will be scored against hidden ground truth using a geometry-focused evaluation metric.

### How It Works

1. **Download** the training dataset (distorted + corrected image pairs) and test images from the Data tab
2. **Build** your lens correction approach — any method you choose
3. **Correct** the test images with your approach and upload them to bounty.autohdr.com
4. **Download** your `submission.csv` from the scoring service
5. **Submit** the CSV here on Kaggle to see your score on the leaderboard

### What Makes a Good Correction?

Our scoring metric prioritizes geometric accuracy:
- **Edge alignment** — do edges in your output match the ground truth?
- **Line straightness** — are lines that should be straight actually straight?
- **Gradient orientation** — are structural directions preserved correctly?
- **Structural similarity** — does the overall structure match?
- **Pixel accuracy** — are the pixels close to ground truth?

You don't need pixel-perfect output. The metric is designed to reward correct geometry even when there are minor differences in color, noise, or compression artifacts.

Any method is fair game. We care about the result, not how you got there.