/MVSS-Net

code for Image Manipulation Detection by Multi-View Multi-Scale Supervision

Primary LanguageJupyter Notebook

MVSS-Net

Code and models for ICCV 2021 paper: Image Manipulation Detection by Multi-View Multi-Scale Supervision

Image text

Something new: MVSS-Net++

We now have an improved version of MVSS-Net, denoted as MVSS-Net++. Check here.

Environment

  • Ubuntu 16.04.6 LTS
  • Python 3.6
  • cuda10.1+cudnn7.6.3

Requirements

Usage

Dataset

An example of the dataset index file is given as data/CASIAv1plus.txt, where each line contains:

img_path mask_path label
  • 0 represents the authentic and 1 represents the manipulated.
  • For an authentic image, the mask_path is "None".
  • For wild images without mask groundtruth, the index should at least contain "img_path" per line.
Training sets
Test sets
  • DEFACTO-12k
  • Columbia
  • COVER
  • NIST16
  • CASIAv1plus: Note that some of the authentic images in CASIAv1 also appear in CASIAv2. With those images fully replaced by Corel images that are new to both CASIAv1 and CASIAv2, we constructed a revision of CASIAv1 termed as CASIAv1plus. We recommend to use CASIAv1plus as an alternative to the original CASIAv1.

Trained Models

We offer FCNs and MVSS-Nets trained on CASIAv2 and DEFACTO_84k, respectively. Please download the models and place them in the ckpt directory:

The performance of these models for image-level manipulation detection (metric: AUC and image-level F1) is as follows. More details are reported in the paper.

Performance metric: AUC
Model Training data CASIAv1plus Columbia COVER DEFACTO-12k
MVSS_Net CASIAv2 0.932 0.980 0.731 0.573
MVSS_Net DEFACTO-84k 0.771 0.563 0.525 0.886
FCN CASIAv2 0.769 0.762 0.541 0.551
FCN DEFACTO-84k 0.629 0.535 0.543 0.840
Performance metric: Image-level F1 (threshold=0.5)
Model Training data CASIAv1plus Columbia COVER DEFACTO-12k
MVSS_Net CASIAv2 0.759 0.802 0.244 0.404
MVSS_Net DEFACTO-84k 0.685 0.353 0.360 0.799
FCN CASIAv2 0.684 0.481 0.180 0.458
FCN DEFACTO-84k 0.561 0.492 0.511 0.709

Inference & Evaluation

You can specify which pre-trained model to use by setting model_path in do_pred_and_eval.sh. Given a test_collection (e.g. CASIAv1plus or DEFACTO12k-test), the prediction maps and evaluation results will be saved under save_dir. The default threshold is set as 0.5.

bash do_pred_and_eval.sh $test_collection
#e.g. bash do_pred_and_eval.sh CASIAv1plus

For inference only, use following command to skip evaluation:

bash do_pred.sh $test_collection
#e.g. bash do_pred.sh CASIAv1plus

Demo

  • demo.ipynb: A step-by-step notebook tutorial showing the usage of a pre-trained model to detect manipulation in a specific image.

Citation

If you find this work useful in your research, please consider citing:

@InProceedings{MVSS_2021ICCV,  
author = {Chen, Xinru and Dong, Chengbo and Ji, Jiaqi and Cao, juan and Li, Xirong},  
title = {Image Manipulation Detection by Multi-View Multi-Scale Supervision},  
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},  
year = {2021}  
}

Acknowledgments

Contact

If you enounter any issue when running the code, please feel free to reach us either by creating a new issue in the github or by emailing