Paper: CVPR2018, Learning Rich Features for Image Manipulation Detection
Code based on Faster-RCNN
This is a rough implementation of the paper. Since I do not have a titan gpu, I made some modifications on the algorithm, but you can easily change them back if you want the exact setting from the paper.
Python 3.6 TensorFlow 1.8.0
- Download vgg16 pre-trained weights from here
- save to /data/imagenet_weights/vgg16.ckpt
- Two-stream neural network model: lib/nets/vgg16.py
- noise stream's weights are randomly initialized
- for accurate prediction, please pre-train noise stream's vgg weights on
ImageNet
and overwrite the trainable setting of noise stream afterSRM
conv layer
- Bounding boxes are predicted by both streams.
- In the paper,
RGB stream
alone predicts bbox more accurately, so you may wanna change that as well (also defined in vgg16.py)
- In the paper,
- Use
main_create_training_set.py
to create training set fromPASCAL VOC
dataset.- The generated dataset will follow the
pascal voc
style, which is also required bytrain.py
- The generated dataset will follow the
Tensorboard
file will be save at/default
- Weights will be save to
/default/DIY_detaset/default
The code requires a large memory GPU. If you do not have a 6G+ GPU, please reduce the number of noise stream conv layers for training.
I will update this repo a few weeks later after I installed the new GPU