# Automatic Lens Correction — Data

来源：https://www.kaggle.com/competitions/automatic-lens-correction/data

本地 Meta Kaggle 快照；不是本次实时页面。

# Dataset Description

## Dataset Description                                                  
                                     
  This competition provides pairs of distorted and corrected lens images  
  for training, and a set of 1,000 distorted test images to correct.


**The photographers who uploaded these images have provided their consent and agreed to the applicable Terms of Service. These images are authorized for use strictly for training purposes and may not be used for any other commercial or non-commercial purposes.**

                                                                          
### Test Images (`test/`)
                                     
  1,000 distorted images that need lens correction. Each file is named
  `{image_id}.jpg`. Your submission should contain corrected versions of
  these images with the same filenames.

### Training Data (`train/`)

  ~23,000 image pairs for developing your correction approach. Each pair
  contains:

  - `original.jpg` — the distorted input image (barrel/pincushion lens
  distortion)
  - `generated.jpg` — the corrected ground truth

  ### Ground Truth (`gt/`)

  The corrected ground truth for the 1,000 test images. These are used by
  the scoring system and are **not available for download**.

  ### How to Submit

  1. Correct all 1,000 test images
  2. Zip your corrected images (named `{image_id}.jpg`)
  3. Upload your zip at [bounty.autohdr.com](https://bounty.autohdr.com)
  to get scored
  4. Download your `submission.csv` and submit it here on Kaggle

  ### Scoring

  Images are scored 0–100 using edge similarity, structural similarity,
  and geometric metrics, calibrated so that submitting uncorrected
  originals scores ~0. Missing images score 0.