Awesome
Semantic-Colorization-GAN
This is a supplementary material for the paper SCGAN: Saliency Map-guided Colorization with Generative Adversarial Network, IEEE Transactions on Circuits and Systems for Video Technology (TCSVT'20).
Arxiv: https://arxiv.org/abs/2011.11377
1 Training
We release the training code in train folder.
The codes require following libs:
- Python=3.6
- PyTorch>=1.0.0
- torchvision>=0.2.1
- Cuda>=8.0
- opencv-python>=4.4.0.46
If you want to train on multispectral data, please refer to the train on multispectral images folder.
The pre-trained global feature network can be found in: https://portland-my.sharepoint.com/:f:/g/personal/yzzhao2-c_my_cityu_edu_hk/ErOcFJc0pilMvCkE53Essi0Bjj89h90l0Y9kEYv390kPEw?e=exNoad
The saliency maps are computed by PFAN: https://github.com/CaitinZhao/cvpr2019_Pyramid-Feature-Attention-Network-for-Saliency-detection
2 Evaluation
Please refer to evaluation folder.
3 Testing Examples
3.1 Colorization Results
We show the representative image of our system.
We provide a lot of results randomly selected from ImageNet and MIT Place 365 validation datasets. These images contain multiple scenes and colors.
3.2 Comparison Results
The comparison results with other fully-automatic algorithms are:
The comparison results with other example-based algorithms are:
3.3 Examples of Semantic Confusion and Object Intervention Problems
We give some examples to illustrate the semantic confusion and object intervention problems intuitively. The SCGAN intergrates the low-level and high-level semantic information and understands how to genrate a reasonable colorization. These settings / architectures help the main colorization network to minimize the semantic confusion and object intervention problems.
There is some examples of semantic confusion problem.
There is some examples of object intervention problem.
To further prove this point, we give more examples about generated attention region and how the saliency map works.
3.4 How our Model Learns at each Epoch
In order to prove our system has the strong fitting ability, we plot the evolution of results of multiple epochs of pre-training term and refinement term. We can see the CNN learns the high-level information at second term.
4 Legacy Image Colorization
4.1 Portrait Photographs
We choose several famous legacy portrait photographs in our experiments. The photographs chosen are with different race, gender, age, and scene. We also select a photo of Andy Lau, which represents the contemporary photographs.
4.2 Landscape Photographs
We choose many landscape photographs by Ansel Adams because the quality is so good. While these photographs are taken from US National Archives (Public Domain).
4.3 Famous Lagacy Photographs
In this section, we select some famous photographs (especially before 1950). And we give a color version of them.
4.4 Other Works
There are many fantastic legacy photography works. Our colorization system still predicts visually high-quality reasonable colorized images.
5 Related Projects
Automatic Colorization: Project Github
Learning Representations for Automatic Colorization: Project Paper Github
Colorful Image Colorization: Project Paper Github
Let there be Color!: Project Paper Github
Deoldify: Project Project2 Github
ColouriseSG: Project
CycleGAN: Project Paper Github
6 Reference
If you think the paper is helpful for your research, please cite:
@article{zhao2020scgan,
title={SCGAN: Saliency Map-guided Colorization with Generative Adversarial Network},
author={Zhao, Yuzhi and Po, Lai-Man and Cheung, Kwok-Wai and Yu, Wing-Yin and Abbas Ur Rehman, Yasar},
journal={IEEE Transactions on Circuits and Systems for Video Technology},
volume={31},
number={8},
pages={3062--3077},
year={2020}
}
7 Find our Latest Works About Image / Video Colorization
A similar work on mobile phone image enhancement is available in this webpage
@inproceedings{zhao2019saliency,
title={Saliency map-aided generative adversarial network for raw to rgb mapping},
author={Zhao, Yuzhi and Po, Lai-Man and Zhang, Tiantian and Liao, Zongbang and Shi, Xiang and others},
booktitle={Proceedings of the International Conference on Computer Vision Workshop},
pages={3449--3457},
year={2019}
}
A SOTA fully-automatic video colorization work is available in this webpage
@article{zhao2022vcgan,
title={VCGAN: Video Colorization with Hybrid Generative Adversarial Network},
author={Zhao, Yuzhi and Po, Lai-Man and Yu, Wing-Yin and Rehman, Yasar Abbas Ur and Liu, Mengyang and Zhang, Yujia and Ou, Weifeng},
journal={IEEE Transactions on Multimedia},
volume={25},
pages={3017-3032},
year={2022}
}
A legacy photo restoration work including scribble-based image colorization is available in this webpage
@inproceedings{zhao2021legacy,
title={Legacy Photo Editing with Learned Noise Prior},
author={Zhao, Yuzhi and Po, Lai-Man and Lin, Tingyu and Wang, Xuehui and Liu, Kangcheng and Zhang, Yujia and Yu, Wing-Yin and Xian, Pengfei and Xiong, Jingjing},
booktitle={Proceedings of the Winter Conference on Applications of Computer Vision},
pages={2103--2112},
year={2021}
}
A SOTA scribble-based video colorization work is available in this webpage
@article{zhao2023svcnet,
title={SVCNet: Scribble-Based Video Colorization Network With Temporal Aggregation},
author={Zhao, Yuzhi and Po, Lai-Man and Liu, Kangcheng and Wang, Xuehui and Yu, Wing-Yin and Xian, Pengfei and Zhang, Yujia and Liu, Mengyang},
journal={IEEE Transactions on Image Processing},
volume={32},
pages={4443-4458},
year={2023}
}