Datasets:
original imagewidth (px) 238 1.28k | tampered imagewidth (px) 238 1.28k | source_corpora stringclasses 6
values | type stringclasses 4
values |
|---|---|---|---|
UCSD 1 | removal | ||
Shanghai | re_o | ||
CUHK | re_o | ||
Shanghai | removal | ||
Shanghai | re_o | ||
Shanghai | removal | ||
Shanghai | removal | ||
Shanghai | re_o | ||
UCSD 2 | removal | ||
UCF | re_o | ||
NTU | addition | ||
UCF | re_t | ||
UCF | re_o | ||
CUHK | re_o | ||
CUHK | re_t | ||
NTU | removal | ||
Shanghai | re_o | ||
Shanghai | addition | ||
UCF | removal | ||
UCF | re_t | ||
CUHK | re_o | ||
UCF | removal | ||
Shanghai | addition | ||
UCSD 1 | removal | ||
UCF | re_o | ||
UCF | re_o | ||
Shanghai | addition | ||
UCF | re_o | ||
UCF | removal | ||
Shanghai | re_t | ||
CUHK | removal | ||
Shanghai | re_o | ||
UCF | re_t | ||
UCF | removal | ||
Shanghai | re_o | ||
Shanghai | re_t | ||
Shanghai | re_o | ||
UCF | re_t | ||
Shanghai | removal | ||
NTU | re_t | ||
UCF | re_t | ||
NTU | removal | ||
UCF | re_t | ||
UCF | re_o | ||
UCSD 1 | removal | ||
Shanghai | addition | ||
NTU | re_o | ||
UCF | re_o | ||
UCF | re_o | ||
Shanghai | re_o | ||
UCF | re_t | ||
NTU | re_t | ||
UCF | removal | ||
UCF | re_o | ||
UCSD 1 | removal | ||
UCSD 1 | re_o | ||
UCF | re_t | ||
UCF | re_o | ||
Shanghai | removal | ||
NTU | re_t | ||
NTU | re_t | ||
UCSD 2 | re_t | ||
Shanghai | addition | ||
UCF | re_t | ||
NTU | re_t | ||
Shanghai | re_t | ||
Shanghai | re_t | ||
UCSD 2 | re_o | ||
UCSD 2 | removal | ||
UCF | removal | ||
UCSD 1 | re_t | ||
UCF | removal | ||
Shanghai | re_o | ||
UCF | re_t | ||
UCF | re_o | ||
Shanghai | re_o | ||
Shanghai | removal | ||
UCSD 1 | removal | ||
NTU | removal | ||
UCSD 1 | removal | ||
UCSD 1 | re_o | ||
Shanghai | re_o | ||
UCF | re_t | ||
Shanghai | re_o | ||
Shanghai | re_t | ||
UCF | re_t | ||
UCF | removal | ||
Shanghai | re_o | ||
UCSD 1 | removal | ||
UCF | re_t | ||
Shanghai | re_o | ||
UCF | removal | ||
NTU | removal | ||
Shanghai | re_o | ||
CUHK | removal | ||
UCF | re_t | ||
UCF | removal | ||
Shanghai | re_t | ||
NTU | re_t | ||
Shanghai | addition |
SurFITR-Detection
Surveillance Forgery Image Test Range (SurFITR) is a dataset for surveillance-style image forgery detection and localisation, published at ACM Multimedia 2026. This repository holds the detection part: real and tampered images with image-level labels and metadata.
- Pixel-level tampering masks are in the gated companion repository wqz995/SurFITR-Localisation.
- Project page and code: github.com/mike-qz-wang/SurFITR
- Paper: ACM DL · arXiv:2604.07101
The dataset viewer above shows a small preview (800 pairs per collection); the full dataset is in the archives below.
Files
| File | Size | Contents |
|---|---|---|
base.tar.gz |
6.3 GB | SurFITR-Base: 108,171 real and 108,171 tampered images, train and test splits, metadata.csv |
transfer.tar.gz |
1.9 GB | SurFITR-Transfer: 29,985 real and 29,985 tampered images, train (Eval 1) and test (Eval 2) splits, metadata.csv |
data/*_preview.parquet |
Viewer previews only |
| Collection | Editing models | Split | Real | Tampered |
|---|---|---|---|---|
| Base | FLUX.1-Fill-dev | train |
52,752 | 52,752 |
test |
55,419 | 55,419 | ||
| Transfer | FLUX.2 klein, Z-Image, Qwen-Image-Edit, Stable Diffusion 3.5 | train (Eval 1) |
14,963 | 14,963 |
test (Eval 2) |
15,022 | 15,022 |
Images come from six surveillance corpora: UCF Crime (1_ucf), NTU Fight (2_ntu), ShanghaiTech (3_shanghai), CUHK Avenue (4_cuhk), UCSD Ped1 (5_ucsd_1) and UCSD Ped2 (6_ucsd_2).
Each tampered image has one of four edit types: addition, removal, re_t (targeted replacement) and re_o (open-ended replacement).
Download
pip install -U huggingface_hub
hf download wqz995/SurFITR-Detection --repo-type dataset --local-dir SurFITR
mkdir -p SurFITR/base SurFITR/transfer
tar -xzf SurFITR/base.tar.gz -C SurFITR/base
tar -xzf SurFITR/transfer.tar.gz -C SurFITR/transfer
Structure
base/ # transfer/ has the same layout
├── metadata.csv
├── real/{1_ucf,...,6_ucsd_2}/{train,test}/00000001.jpg
└── fake/{1_ucf,...,6_ucsd_2}/{train,test}/00000001.jpg
metadata.csv has one row per image: filename, split, source_dataset, label (real / fake), editing_type, and for Transfer also model.
Masks from SurFITR-Localisation extract into the same folders as 00000001_mask.jpg next to each tampered image.
License
CC BY-NC 4.0, for research use only. Source images come from publicly available surveillance datasets; please also respect their terms.
Citation
If you find SurFITR useful for your research, please cite our paper.
@inproceedings{wang2026surfitr,
author = {Wang, Qizhou and Pang, Guansong and Leckie, Christopher},
title = {SurFITR: A Dataset for Surveillance Image Forgery Detection and Localisation},
booktitle = {Proceedings of the 34th ACM International Conference on Multimedia},
series = {MM '26},
year = {2026},
pages = {16972--16979},
doi = {10.1145/3767308.3838668},
url = {https://doi.org/10.1145/3767308.3838668}
}
- Downloads last month
- 81