FGRestore Dataset README

This directory contains the FGRestore data and annotations for two tasks:

• Image quality score prediction (Regression)
• Pairwise preference prediction (Rank)

Directory Layout

• annotations/IR_train.json: training set (contains both score and preference annotations)
• annotations/test_single/: test split for score prediction
• annotations/test_pairs/: test split for pairwise preference prediction
• images/: distorted or restored images
• ref/: reference images

Annotation Formats

1) Training: annotations/IR_train.json

Each sample is a pairwise record with both preference and score labels. Typical fields:

• ref_image: reference image path relative to FGRestore (for example, ref/000000.png)
• image_nameA: candidate image A path (for example, images/000000.png)
• image_nameB: candidate image B path (for example, images/000001.png)
• human_preference_for_A: human preference label
  • 1: A is preferred over B
  • 0: B is preferred over A
  • 0.5: tie between A and B
• score_normA: normalized score of image A
• score_normB: normalized score of image B
• task: restoration task name (for example, Denoising)
• scene_id: scene identifier

2) Score Test Split: annotations/test_single/

Used to evaluate score prediction performance. Each record is a single-image sample. Typical fields:

• ref_image
• image_name
• score_norm
• task
• scene_id

3) Preference Test Split: annotations/test_pairs/

Used to evaluate pairwise preference performance. Each record is a pairwise sample. Typical fields:

• ref_image
• image_nameA
• image_nameB
• human_preference_for_A (can be 0, 0.5, or 1)
• score_normA
• score_normB
• task
• scene_id

Official Train/Test Usage

• Training: use FGRestore/annotations/IR_train.json
• Score evaluation: use FGRestore/annotations/test_single
• Preference evaluation: use FGRestore/annotations/test_pairs

Important Note for Scene IDs 3 and 4

For samples with scene_id equal to 3 or 4, a true pristine reference image does not exist.

In these cases, ref_image corresponds to the pre-restoration image, not a distortion-free ground-truth reference.

Suggested Metrics

• Score prediction (test_single): SRCC / PLCC
• Preference prediction (test_pairs): Accuracy

Data Loading Notes

• All path fields are relative paths; use FGRestore as the data root when resolving files.