KhangTruong/MVSS-train-set

This repository provides standardized, manipulation-only benchmark training and validation sets established under the MVSS Protocol (introduced by Dong et al. in MVSS-Net: Image Manipulation Detection by Multi-View Multi-Scale Supervision, ICCV 2021 & IEEE TPAMI 2022).

All data files reside in a unified directory structure under data/, partitioned into compact ~200 MB Parquet files for fast, memory-efficient streaming and multi-worker distributed training.

---

1. Key Highlights

1. Manipulation/Forgery-Only (Pure Fakes):

   • Authentic / negative images (solid black masks with no tampered pixels) have been segregated.
   • 100% of samples in this repository are verified forgeries where the ground truth binary mask contains positive white pixels (> 0) denoting modified/tampered visual content.

2. Unified Single-Directory Layout:

   • Both CASIA v2.0 and DEFACTO-84k files reside together in the data/ folder, distinguishable by intuitive prefix patterns:
     • data/train-casia-v2.0-*.parquet
     • data/train-defacto-84k-*.parquet
     • data/validation-casia-v2.0-*.parquet
     • data/validation-defacto-84k-*.parquet

3. Zero-Configuration Loading:

   • Calling load_dataset("KhangTruong/MVSS-train-set", split="train") automatically loads both benchmark subsets concurrently as a single combined training set without needing to specify a configuration subset name.
   • Individual sub-benchmarks can still be loaded selectively via config_name="casia-v2.0" or config_name="defacto-84k" (or aliases casia, defacto).

4. Optimal Shard Sizing (~200 MB):

   • Parquet shards are consistently partitioned to ~200 MB each, preventing out-of-memory errors and facilitating seamless multi-worker DataLoader prefetching.

---

2. Benchmark Overview

Configuration	Split	Examples	Forgery Categories	Shard File Pattern	Shards & Approx Size
casia-v2.0	train	4,354	Splicing (Tp_D_*), Copy-Move (Tp_S_*)	data/train-casia-v2.0-*.parquet	12 shards (~200 MB each)
casia-v2.0	validation	769	Splicing (Tp_D_*), Copy-Move (Tp_S_*)	data/validation-casia-v2.0-*.parquet	2 shards (~200 MB each)
defacto-84k	train	64,409	Splicing (34,133), Inpainting (17,507), Copy-Move (12,769)	data/train-defacto-84k-*.parquet	166 shards (~198 MB each)
defacto-84k	validation	16,088	Splicing (7,179), Inpainting (5,494), Copy-Move (3,415)	data/validation-defacto-84k-*.parquet	42 shards (~199 MB each)
default (Unified)	train	68,763	All CASIA + DEFACTO manipulation training samples	data/train-*	178 shards (~200 MB each)
default (Unified)	validation	16,857	All CASIA + DEFACTO manipulation validation samples	data/validation-*	44 shards (~200 MB each)

---

3. Dataset Schema

Every sample across all files adheres to the standardized schema:

Column	Type	Description
image	datasets.Image	The tampered image in RGB format (JPEG / PNG / TIFF).
mask	datasets.Image	Pixel-level binary manipulation mask (single-channel L mode). 0 (Black) indicates authentic background, 255 (White) indicates forged/modified pixels.

---

4. Usage Examples

A. Load the Unified Training Set (No Subset Specification Needed)

from datasets import load_dataset

# Automatically loads both CASIA-v2.0 and DEFACTO-84k fakes combined
train_set = load_dataset("KhangTruong/MVSS-train-set", split="train")
val_set = load_dataset("KhangTruong/MVSS-train-set", split="validation")

print(f"Total unified train: {len(train_set)} samples")
print(f"Total unified validation: {len(val_set)} samples")

# Inspect a sample
sample = train_set[0]
image = sample["image"]  # PIL.Image RGB
mask = sample["mask"]    # PIL.Image L (contains white pixels > 0)

B. Load Specific Subsets Individually

# 1. Load CASIA v2.0 only
casia_train = load_dataset("KhangTruong/MVSS-train-set", "casia-v2.0", split="train")
casia_val = load_dataset("KhangTruong/MVSS-train-set", "casia-v2.0", split="validation")
print(f"CASIA v2.0: {len(casia_train)} train, {len(casia_val)} val")

# 2. Load DEFACTO-84k only (Forgery-Only)
defacto_train = load_dataset("KhangTruong/MVSS-train-set", "defacto-84k", split="train")
defacto_val = load_dataset("KhangTruong/MVSS-train-set", "defacto-84k", split="validation")
print(f"DEFACTO-84k: {len(defacto_train)} train, {len(defacto_val)} val")

C. Low-Memory Streaming

# Stream samples without downloading full dataset locally
stream_train = load_dataset("KhangTruong/MVSS-train-set", split="train", streaming=True)
for sample in stream_train:
    img = sample["image"]
    mask = sample["mask"]
    break

---

5. Citations

If you use this benchmark dataset in your research, please cite:

@inproceedings{dong2021mvss,
  title={MVSS-Net: Multi-View Multi-Scale Supervision for Image Manipulation Detection},
  author={Dong, Xiaodong and Bao, Jing and Chen, Dongdong and Zou, Weiming and Zhang, Hao and Fang, Lu and Yuan, Lu and Chen, Dong},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2021}
}

@article{dong2022mvss,
  title={MVSS-Net: Image Manipulation Detection by Multi-View Multi-Scale Supervision},
  author={Dong, Xiaodong and Bao, Jing and Chen, Dongdong and Zou, Weiming and Zhang, Hao and Fang, Lu and Yuan, Lu and Chen, Dong},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
  year={2022}
}

@inproceedings{mahfoudi2019defacto,
  title={DEFACTO: Image and Face Manipulation Dataset},
  author={Mahfoudi, Ga{\"e}l and Tajini, Roman and Retraint, Florence and Blaise, Fr{\'e}d{\'e}ric and Pic, Marc and Combier, Nathalie},
  booktitle={2019 27th European Signal Processing Conference (EUSIPCO)},
  year={2019}
}

@inproceedings{dong2013casia,
  title={CASIA Image Tampering Detection Evaluation Database},
  author={Dong, Jing and Wang, Wei and Tan, Tieniu},
  booktitle={2013 IEEE China Summit and International Conference on Signal and Information Processing},
  year={2013}
}