wasidennis / deepharmonization Goto Github PK

Demo code of the paper: "Deep Image Harmonization", Y.-H. Tsai, X. Shen, Z. Lin, K. Sunkavalli, X. Lu and M.-H. Yang, CVPR 2017

Python 100.00%

computer-vision computational-photography image-editing deep-learning encoder-decoder

deepharmonization's Introduction

Deep Image Harmonization

Project webpage: https://sites.google.com/site/yihsuantsai/research/cvpr17-harmonization
Contact: Yi-Hsuan Tsai (wasidennis at gmail dot com)

Paper

Deep Image Harmonization
Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu and Ming-Hsuan Yang
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017.

This is the authors' demo code described in the above paper. Please cite our paper if you find it useful for your research.

One re-implementation of our dataset and model: https://github.com/bcmi/Image_Harmonization_Datasets

Installation and Usage

Download and unzip the code.
Install Caffe: http://caffe.berkeleyvision.org/.
Download the pre-trained caffe model and move it under the model folder.
Run demo.py on real composite images (including our test set collected in the paper).

Evaluation Set

Download our complete set of real composite images, including our harmonization results here.

Note

The model, code and dataset are available for non-commercial research purposes only.

Log

03/2017: demo code released
05/2017: complete evaluation set released

deepharmonization's People

Contributors

Stargazers

Watchers

deepharmonization's Issues

Save as a large size image

Is there a way to save as a 1024x1024px image?

Curated Dataset for Training

Is it possible for you to share the complete dataset (including ground truth) you used for training the model (apart from test images you shared)?

Hi,Would you mind release the training code?

Thank you!

About your network achitecture

Hi, Thanks for your useful code, I have a problem when I trying to reproduce your experiment.
For the last Conv layer used in the network, you said it's a 4x4 conv layer in your paper, but in your prototxt, it's a fully-connected layer. If you could answer my question, It would be very appreciated.

Running demo.py resulting in complete white images

Hi,
I just tried to run the demo.py script, and the results are all white. I wanted to debug it to see the output values before the clipping to 0,255 but I don't see what I'm trying to print (I'm new to caffe - is it not possible to print to std out on inference?).

Caffe version: 0.17.0
Ubuntu: 16.04.2
Python: 2.7.12

Thank you!
Adva

Can not initiate a model

Hi guys,
Thanks for the interesting paper and released code!

I've downloaded a model (harmonize_iter_200000.caffemodel) but it's not working:

root@6400dcb7b629:/DeepHarmonization# python demo.py
...
I0222 20:06:22.437355    17 net.cpp:284] Network initialization done.
I0222 20:06:22.616809    17 net.cpp:791] Ignoring source layer data
I0222 20:06:22.616827    17 net.cpp:791] Ignoring source layer mask_data_1_split
F0222 20:06:22.616842    17 net.cpp:797] Check failed: target_blobs.size() == source_layer.blobs_size() (5 vs. 3) Incompatible number of blobs for layer bn0
*** Check failure stack trace: ***
Aborted (core dumped)

Are there any specific requirements I should care about?

I wonder how to use adobe 5k datasets?

Is there any semantic information provided in this dataset?

Real Composite Images.

Excuse me, could you please provide the DIH99 dataset？