README.md

metadata

pretty_name: >-
  Enhanced Parquet Edition: IMD2020: A Large-Scale Annotated Dataset Tailored
  for Detecting Manipulated Images
language:
  - en
tags:
  - parquet
  - image
  - image-forgery
  - image-manipulation-detection
size_categories:
  - 1K<n<10K
task_categories:
  - image-classification
  - image-segmentation
dataset_info:
  - config_name: imd2020
    features:
      - name: image
        dtype: image
      - name: mask
        dtype: image
      - name: sample_id
        dtype: string
      - name: label
        dtype: int64
      - name: manipulation
        dtype: string
      - name: source_split
        dtype: string
      - name: filename_source_id
        dtype: string
      - name: source_image_path
        dtype: string
      - name: source_mask_path
        dtype: string
      - name: width
        dtype: int64
      - name: height
        dtype: int64
      - name: has_mask
        dtype: bool
    splits:
      - name: full
        num_examples: 2424
configs:
  - config_name: imd2020
    default: true
    data_files:
      - split: full
        path:
          - imd2020/parquet/full-00000-of-00008.parquet
          - imd2020/parquet/full-00001-of-00008.parquet
          - imd2020/parquet/full-00002-of-00008.parquet
          - imd2020/parquet/full-00003-of-00008.parquet
          - imd2020/parquet/full-00004-of-00008.parquet
          - imd2020/parquet/full-00005-of-00008.parquet
          - imd2020/parquet/full-00006-of-00008.parquet
          - imd2020/parquet/full-00007-of-00008.parquet

Enhanced Parquet Edition: IMD2020: A Large-Scale Annotated Dataset Tailored for Detecting Manipulated Images

This repository packages the IMD2020 archives distributed by ForgeryVCR into a Hugging Face Parquet configuration. It contains 2,424 images (414 real and 2,010 tampered), with image bytes, supplied masks, and per-sample metadata embedded in Parquet.

The source archives are the ForgeryVCR 832-size preprocessed release under datasets/train/rl. The full split contains all supplied samples and introduces no new train/test partition.

Subsets

Config	Split	Rows	Image size	Pixel masks	Parquet files
imd2020	full	2,424	832 × 832	Tampered images	8

Columns

• image: the image from the supplied archive, decoded with the Hugging Face Image feature.
• mask: the supplied grayscale annotation; white marks edited regions and black marks background. Intermediate grayscale values are preserved. Real images have a null mask.
• label: image-level authenticity label, 0 = real, 1 = fake.
• manipulation: real or tampered.
• sample_id, source_image_path: the image path relative to the extracted release directory.
• source_split: the upstream ForgeryVCR directory, train/rl.
• filename_source_id: the original filename stem.
• source_mask_path: the mask path relative to the extracted release directory; null for real images.
• width, height: supplied image dimensions in pixels.
• has_mask: whether a pixel annotation is supplied.

Metadata Files

Each configuration includes stats.json with sample and label counts, image dimensions, mask statistics, and Parquet shard paths and sizes.

The root inventory.json collects these statistics and records the total row count, shard count, and encoded byte sizes. provenance/ retains the supplied sample listings and bounding-box annotations.

Citation

If you reference IMD2020, please cite the original paper:

@inproceedings{Novozamsky_2020_WACV,
  author={Novozamsky, Adam and Mahdian, Babak and Saic, Stanislav},
  title={IMD2020: A Large-Scale Annotated Dataset Tailored for Detecting Manipulated Images},
  booktitle={2020 IEEE Winter Applications of Computer Vision Workshops (WACVW)},
  pages={71--80},
  year={2020},
  doi={10.1109/WACVW50321.2020.9096940}
}

Dataset Export

This Parquet edition reorganizes the ForgeryVCR IMD2020 release and adds per-sample metadata. Supplied image and mask bytes are preserved. Please also reference this dataset repository when using the packaged version.

@misc{thanejoss_imd2020_2026,
  author={{ThaneJoss}},
  title={{IMD2020}},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/ThaneJoss/IMD2020},
  note={Dataset available on Hugging Face}
}