Skip to content

Latest commit

 

History

History
72 lines (53 loc) · 3.25 KB

README_EN.md

File metadata and controls

72 lines (53 loc) · 3.25 KB
Raw

PhotoWCT

Unofficial implementation of "A Closed-form Solution to Photorealistic Image Stylization"

中文介绍

Brief introduction

This Project is the Unofficial implementation of "A Closed-form Solution to Photorealistic Image Stylization" which is a fork from the project https://github.com/eridgd/WCT-TF

The official implementation is located in https://github.com/NVIDIA/FastPhotoStyle and this project follow the TUTORIAL.md Example 3 steps in it.

Paper Link: https://arxiv.org/abs/1802.06474

Construction Steps

Data Prepare

  • 1 download COCO dataset from http://cocodataset.org/ (this use coco 2017) download vgg_normalized.t7 from web (download_vgg.sh)/

  • 2 This step may not be necessary if you not want to use segmentation as prior in input (with may require more times in wavelet pooling setting, set use_wavelet_pooling = True in WCTModel construction).

    use project https://github.com/CSAILVision/semantic-segmentation-pytorch to generate some segementations
    for coco dataset, if you have some confusions of use weights, you can appeal to my issue comment in CSAILVision/semantic-segmentation-pytorch#116 Then you can edit dataloader in test.py to generate your own segmentations, and save them in "path/to/val2017".replace("2017", "2017_seg") or "path/to/train2017".replace("2017", "2017_seg")

  • 3 Use merge_pics_segs/merge.py to merge pics and segmentations, save them in "path/to/val2017".replace("2017", "2017_with_label") or "path/to/train2017".replace("2017", "2017_with_label")

Train and Evaluate step

The merged conclusion can seen in path/to/simple2017/style_with_label.png or path/to/simple2017/content_with_label.png

  • 4 Train model: 

    python train_edit.py --relu-target relu3_1 --content-path path/to/train2017_with_label --batch-size 8 --feature-weight 1 --pixel-weight 1 --tv-weight 0.0 --learning-rate 1e-4 --max-iter 50000 --val-path path/to/val2017_with_label --checkpoint path/to/ckpt/

  • 5 Evaluate the conclusion:

    python stylize.py --input-checkpoint C:\Coding\Python\PhotoWCT\ckpt\model.ckpt-4601 --relu-targets relu3_1 --alpha 0.8 --style-path path/to/simple2017/style_with_label.png --content-path path/to/simple2017/content_with_label.png --out-path C:\Coding\Python\WCT_tf_final\output --passes 1

  • 6 Apply Haze Removal to conclusion, online in http://tools.pfchai.com
    or with the script located in https://github.com/pfchai/Haze-Removal/blob/master/HazeRemovalWidthGuided.py

Contact

svjack - [email protected] - [email protected]

Project Link:https://github.com/svjack/PhotoWCT