README.md

---
pretty_name: DPDD
license: mit
configs:
- config_name: combined
  data_files:
  - split: train
    path: combined/train-*
  - split: val
    path: combined/val-*
  - split: test
    path: combined/test-*
- config_name: left
  data_files:
  - split: train
    path: left/train-*
  - split: val
    path: left/val-*
  - split: test
    path: left/test-*
- config_name: right
  data_files:
  - split: train
    path: right/train-*
  - split: val
    path: right/val-*
  - split: test
    path: right/test-*
dataset_info:
- config_name: combined
  features:
  - name: source
    dtype: image
  - name: target
    dtype: image
  splits:
  - name: train
    num_bytes: 5955996924.0
    num_examples: 350
  - name: val
    num_bytes: 1287133812.0
    num_examples: 74
  - name: test
    num_bytes: 1286755475.0
    num_examples: 76
  download_size: 8530312203
  dataset_size: 8529886211.0
- config_name: left
  features:
  - name: source
    dtype: image
  splits:
  - name: train
    num_bytes: 2934714938.0
    num_examples: 350
  - name: val
    num_bytes: 631641910.0
    num_examples: 74
  - name: test
    num_bytes: 635338302.0
    num_examples: 76
  download_size: 4201900846
  dataset_size: 4201695150.0
- config_name: right
  features:
  - name: source
    dtype: image
  splits:
  - name: train
    num_bytes: 2938871158.0
    num_examples: 350
  - name: val
    num_bytes: 632939190.0
    num_examples: 74
  - name: test
    num_bytes: 636422517.0
    num_examples: 76
  download_size: 4208438874
  dataset_size: 4208232865.0
---

# DPDD

The DPDD dataset (Dual-Pixel Defocus Deblurring) provides paired defocused and all-in-focus images, along with dual-pixel sub-aperture views, for training and evaluating models on defocus-deblurring tasks. Each scene includes a defocused image captured with a wide aperture, its left/right dual-pixel views, and a corresponding sharp image captured with a small aperture.

This is a Hugging Face compatible version created from the original dataset released with the paper **Defocus Deblurring Using Dual-Pixel Data (ECCV 2020)** by Abdullah Abuolaim and Michael S. Brown.

The original repository is licensed under the **MIT License**, and so does this version. When using the dataset, please cite the original paper as requested by the authors.

## Usage

```py
from datasets import load_dataset

ds = load_dataset("JacobLinCool/DPDD")
print(ds)
# DatasetDict({
#     train: Dataset({
#         features: ['source', 'target'],
#         num_rows: 350
#     })
#     val: Dataset({
#         features: ['source', 'target'],
#         num_rows: 74
#     })
#     test: Dataset({
#         features: ['source', 'target'],
#         num_rows: 76
#     })
# })

print(ds["train"][0])
# {'source': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=1680x1120>,
#  'target': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=1680x1120>}
```