RGB-N

Code and synthetic dataset generation for the CVPR 2018 paper "Learning Rich Features for Image Manipulation Detection"

Environment

tensorflow 0.12.1, python3.5.2, cuda 8.0.44 cudnn 5.1

Other packages please run:

pip install -r requirements.txt

Compile lib and compact_bilinear_pooling:

Check if the cuda lib path in compact_bilinear_pooling/sequential_fft/complie.sh is correct.
Run the command:

cd lib
make
cd compact_bilinear_pooling/sequential_fft
./compile.sh

For more detail, see https://github.com/ronghanghu/tensorflow_compact_bilinear_pooling

Pre-trained model

For ImageNet resnet101 pre-trained model, please download from https://github.com/endernewton/tf-faster-rcnn

Synthetic dataset

Download COCO 2014 dataset (http://cocodataset.org/#download) and COCO PythonAPI (https://github.com/cocodataset/cocoapi) and put in coco_synthetic folder. After this step the coco dataset folder 'cocostuff' will be created.
Change dataDir in coco_synthetic/demo.py to the path of 'train2014' (e.g, ./cocostuff/coco/train2014)
Run run_demo.sh 1 100 to choose the begin and end COCO category used for creating the tamper synthetic dataset.
Run split_train_test.py to make train/test split. (making sure that the images used to generate training set not overlap with the images for testing)

Train on synthetic dataset

Change the coco synthetic path in lib/factory.py:

coco_path= #FIXME
for split in ['coco_train_filter_single', 'coco_test_filter_single']:
    name = split
    __sets[name] = (lambda split=split: coco(split,2007,coco_path))

Specify the ImageNet resnet101 pretrain model path in train_faster_rcnn.sh as below:

        python3 ./tools/trainval_net.py \
            --weight /path of res101.ckpt/data/imagenet_weights/res101.ckpt \ #FIXME
            --imdb ${TRAIN_IMDB} \
            --imdbval ${TEST_IMDB} \
            --iters ${ITERS} \
            --cfg cfgs/${NET}.yml \
            --net ${NET} \
            --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} TRAIN.STEPSIZE ${STEPSIZE} ${EXTRA_ARGS}

Specify the dataset, gpu, and network in train_dist_faster.sh as below as run the file

./train_faster_rcnn.sh 0 coco res101_fusion EXP_DIR coco_flip_0001_bilinear_new

Use synthetic pre-trained model for fine tuning

Specify the ImageNet resnet101 pretrain model path in train_faster_rcnn.sh as below:

        python3 ./tools/trainval_net.py \
            --weight /path of synthetic pretrain model/res101_fusion_faster_rcnn_iter_60000.ckpt \ #FIXME
            --imdb ${TRAIN_IMDB} \
            --imdbval ${TEST_IMDB} \
            --iters ${ITERS} \
            --cfg cfgs/${NET}.yml \
            --net ${NET} \
            --set ANCHOR_SCALES ${ANCHORS} ANCHOR_RATIOS ${RATIOS} TRAIN.STEPSIZE ${STEPSIZE} ${EXTRA_ARGS}

Specify the dataset, gpu, and network in train_dist_faster.sh as below as run the file (use NIST as an example)

./train_faster_rcnn.sh 0 NIST res101_fusion EXP_DIR NIST_flip_0001_bilinear_new

Test the model

Check the model path match well with NET_FINAL in test_faster_rcnn.sh, making sure the checkpoint iteration exist in model output path. Otherwise, change the iteration number ITERS as needed.

  coco)
    TRAIN_IMDB="coco_train_filter_single"
    TEST_IMDB="coco_test_filter_single"
    ITERS=60000
    ANCHORS="[8,16,32,64]"
    RATIOS="[0.5,1,2]"
    ;;

Run test_dist_faster.sh. If things go correcty, it should print out MAP and save tamper.txt and tamper.png indicating the detection result and PR curve.

Synthetic dataset and training/testing split

https://drive.google.com/open?id=1vIAFsftjmHg2J5lJgO92C1Xmyw539p_B

Citation:

If this code or dataset helps your research, please cite our paper:

@inproceedings{zhou2018learning,
  title={Learning Rich Features for Image Manipulation Detection},
  author={Zhou, Peng and Han, Xintong and Morariu, Vlad I and Davis, Larry S},
  booktitle = {CVPR},
  year={2018}
}

dokooh/RGB-N