Myung-Joon Kwon

I've uploaded the weights for RGB Stream Only (RGB_only_v2.tar) to Google Drive. Thanks

Okay, I was not meant to upload RGB Stream code... but since you requested I'm posting it here. You should create splicing_dataset something like: `splicing_dataset(crop_size=None, grid_crop=True, blocks=('RGB',), DCT_channels=1, mode='arbitrary', read_from_jpeg=True)`...

Hi, CAT-Net is the first work to use DCT coefficients in CNN for forgery localization. While there have been follow-up works that have used DCT coefficients, as far as I...

Are you talking about training? In training, auth and tamp are not used as pairs. They are used independently. You can use any auth and tamp images. You may even...

1. If tamp and gt have different sizes, gt is resized to match the tamp size. 2. I cannot understand your sentences.

I think the most critical problem is the lack of training data. I recommend you to collect more. Also, why don’t you add momentum in sgd?

@huytuong010101 Here's the plot :grinning: 

Hi, I chose the class weights by the number of authentic and tampered pixels throughout the datasets. To be specific, # auth pixels : # tamp pixels = tamp class...

Okay, I would like to support your work of implementing CAT-Net with Tensorflow. I will upload the tampCOCO and JPEG RAISE datasets tomorrow to Baidu cloud. I cannot read Chinese...

@huytuong010101 I've just uploaded all custom datasets used in the paper to Google Drive. The datasets exceeded 105GB so I couldn't upload them to Baiduyun. Fortunately, I found that my...