File size: 6,720 Bytes
24f1bad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
---
license: cc-by-4.0
task_categories:
- image-segmentation
- image-classification
language:
- en
tags:
- image-forensics
- copy-move-forgery
- forgery-detection
- tamper-detection
- synthetic
- coco
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---

# COCO-CMFD

A synthetic copy-move forgery dataset generated from MS-COCO 2017, with
source/target-separated ground truth for copy-move forgery detection
(CMFD).

Each sample takes one annotated COCO object, applies a mild affine
transform, and pastes it elsewhere in the *same* image at a location
that passes scene-plausibility checks (support surface, horizon band,
perspective scale, occlusion). Ground truth is provided as a 3-class
trimap, a binary mask, a 16 px patch-label grid, and a hard-negative
mask of the scene's other objects.

Code, generation pipeline and quality tooling:
https://github.com/harshitajain523/COCO-Copy-Move-Forgery-Dataset

## Dataset details

| | |
|---|---|
| Samples | 12,271 |
| Source images | 7,150 distinct COCO train2017 images (1–8 samples each, mean 1.7) |
| Categories | 64 COCO categories across 9 supercategories |
| Resolution | native COCO, 320–640 px per side |
| Pasted region size | 2.5%–47.6% of image area (median 8.6%) |
| Transforms | scale 0.70–1.45, rotation ±10°, horizontal flip on 25.1% |
| Post-processing | none (no JPEG, noise or blur applied) |

### Fields

| Field | Type | Description |
|---|---|---|
| `image` | Image | Forged image (PNG, lossless) |
| `trimap` | Image | `0` background, `128` source, `255` target |
| `binary_mask` | Image | `0`/`255`, source ∪ target |
| `hard_negatives` | Image | `0`/`255`, all non-source annotated objects |
| `patch_labels` | list[int8] | Flattened 16 px grid: `-1` ignore, `0` bg, `1` source, `2` target |
| `patch_labels_shape` | list[int] | `[H/16, W/16]` for reshaping `patch_labels` |
| `gate_profile` | string | `strict`, `relaxed_v1`, or `relaxed_v2` (see below) |
| `source_reused` | int | `1` if this source object is also the source of another sample from the same image |
| `source_category_name` | string | COCO category of the copied object |
| `scale_factor`, `rotation_angle`, `flip` | float/int | Transform applied to the copy |
| `source_bbox`, `target_bbox` | string | `(x, y, w, h)` tuples |
| `placement_tier`, `horizon_tier` | string | Which candidate pool / vertical constraint accepted the placement |

24 metadata fields in total; `image_id` refers to the originating COCO
image.

Patch labels use a strict purity rule: a patch is labelled only if all
256 pixels share one class, otherwise `-1`. This keeps the copy-move
boundary out of patch-level contrastive objectives.

### Generation profiles

| `gate_profile` | Samples | Source-object gates |
|---|---|---|
| `strict` | 4,420 | edge margin 15 px, isolation 5×5, bbox fill ≥ 0.25, paste ≥ 2% of image |
| `relaxed_v1` | 166 | edge 8 px, isolation 3×3, fill ≥ 0.20 |
| `relaxed_v2` | 7,685 | as `relaxed_v1`, plus paste ≥ 1.2% of image |

Scene-plausibility gates are identical across all three profiles; only
source-object selectivity differs. Filter on `gate_profile == "strict"`
for the most conservative subset.

## Usage

```python
from datasets import load_dataset
import numpy as np

ds = load_dataset("harshitajainn/coco-cmfd", split="train")
s = ds[0]

image = s["image"]                              # PIL.Image
trimap = np.array(s["trimap"])
source_mask, target_mask = trimap == 128, trimap == 255
patches = np.array(s["patch_labels"], dtype=np.int8).reshape(
    s["patch_labels_shape"]
)
```

WebDataset tar shards are also published under `wds/` for
streaming-heavy pipelines; each sample key carries `.png`,
`.trimap.png`, `.binary.png`, `.hardneg.npy`, `.patches.npy` and
`.json`.

## Verification

Run on the released set with the tooling in the GitHub repository:

- **Mask alignment** (60 samples): IoU between changed pixels and the
  labelled target region — mean 1.0000, min 0.9997, with 0 pixels
  modified outside any mask. Blending uses an inward-only 2 px feather,
  so the alpha ramp stays inside the labelled region and the ground
  truth is exact rather than approximate.
- **Integrity**: 61,355/61,355 referenced files present, 400 sampled
  pixel checks pass.
- **Post-hoc plausibility audit**: 12,190 clean / 81 rejected (0.66%).

## Intended use and limitations

Intended for pretraining and evaluating copy-move forgery detectors,
including source/target discrimination and patch-level contrastive
objectives.

Limitations:

- **Synthetic**. Forgeries are algorithmically composited, not made by
  a human with intent to deceive. Detectors trained only on this data
  should be fine-tuned and evaluated on real-world benchmarks.
- **No post-processing**. Images are lossless PNG with no JPEG
  recompression, noise or blur. Robustness to those degradations is not
  exercised by this set.
- **Mild transforms only** (±10° rotation, 0.70–1.45× scale). Large
  rotations and extreme rescaling are out of distribution.
- **Category imbalance** follows COCO: person 8.1%, bird 7.5%, clock
  7.2% are the most frequent of 64 categories.
- **Source correlation**: 4,050 samples reuse a source object that
  another sample from the same image also uses. Group by `image_id`
  when constructing splits to avoid leakage.
- **Single copy-move per image**; no multi-source or nested forgeries.

## License and attribution

Released under **CC BY 4.0**.

Images derive from MS-COCO 2017. COCO *annotations* are CC BY 4.0;
COCO *images* originate from Flickr and remain subject to their owners'
terms — the COCO Consortium does not hold their copyright. The CC BY
4.0 grant covers the forgery generation, ground-truth masks and
metadata, not the underlying photographic content. Review the
[COCO terms of use](https://cocodataset.org/#termsofuse) before
redistributing the imagery, particularly for commercial use.

## Citation

```bibtex
@dataset{jain_coco_cmfd_2026,
  author    = {Jain, Harshita},
  title     = {{COCO-CMFD}: A Synthetic Copy-Move Forgery Dataset
               with Source/Target Ground Truth},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {TODO},
  url       = {https://github.com/harshitajain523/COCO-Copy-Move-Forgery-Dataset}
}
```

Please also cite MS-COCO:

```bibtex
@inproceedings{lin2014microsoft,
  title     = {Microsoft {COCO}: Common Objects in Context},
  author    = {Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and
               Hays, James and Perona, Pietro and Ramanan, Deva and
               Doll{\'a}r, Piotr and Zitnick, C Lawrence},
  booktitle = {ECCV},
  year      = {2014}
}
```