Fashion++: Minimal Edits for Outfit Improvement

Wei-Lin Hsiao, Isay Katsman*, Chao-Yuan Wu*, Devi Parikh, Kristen Grauman
In ICCV 2019. [paper]

Installation

This project is tested on Ubuntu 18.04, with python 3.6.4, pytorch 0.4

• Install pytorch >= 0.4 following instructions from https://pytorch.org/
• Install python library dominate

pip install dominate

• Install python library opencv

pip install opencv-python

• Clone this repo:

git clone https://github.com/facebookresearch/FashionPlus.git

Demo

1. Download pretrained models from:

   Model	file name	link
   cGAN	latest_net_E.pth	model
   cGAN	latest_net_G.pth	model
   cGAN	latest_net_D.pth	model
   cGAN	train_avg_features.p	features
   VAE	latest_Decoder.pth	model
   VAE	latest_Separate_encoder.pth	model
   VAE	latest_Together_encoder.pth	model
   classifier	120_classifier.pth	model

   Place all but 120_classifer.pth in checkpoint/humanparsing/.
   Place 120_classifier.pth in checkpoint/.

2. Prepare images in datasets/images/.
   Prepare corresponding segmentation masks in datasets/labels/.
   Note: Our pretrained models depend on the label taxonomy it was trained on, which is from: HumanParsing-Dataset

# label index:  segment name
0            :  null
1            :  hat
2            :  hair
3            :  glasses
4            :  up(per clothing)
5            :  skirt
6            :  pants
7            :  dress
8            :  belt
9            :  r-shoe
10           :  l-shoe
11           :  face
12           :  r-leg
13           :  l-leg
14           :  r-arm
15           :  l-arm
16           :  bag
17           :  scarf

Segmentation labels need to be converted into the above mapping in order to work with FashionPlus.

3. Prepare input data for models:
   Change ROOT_DIR in script to FashionPlus' absolute path on your system.

cd preprocess
./run_prepare_data.sh

4. Encode input images into latent codes:
   Change ROOT_DIR in script to FashionPlus' absolute path on your system.

cd preprocess
./encode_shape_texture_features.sh

5. Perform edit and generate the changed output:
   Change ROOT_DIR in script to FashionPlus' absolute path on your system.
   Specify options:

   • UPDATE_FNAME: image file's name to be edited
   • UPDATE_TYPE: could choose to update only shape, or only texture, or both
   • AUTO_SWAP: whether to let the editing module automatically decide which garment to change; if specified, SWAPPED_PARTID will have no effect
   • SWAPPED_PARTID: explicitly specify which garment to swap out; here our mapping is top=0, skirt=1, pants=2, dress=3
   • MAX_ITER: if the updating module's stop criterion is among reaching the maximal iteration, then specify the maximal value here
   • STEP_SZ: amount of edit performed in each iteration; the larger the step size, the more significant the change is

./scripts/edit_and_visualize_demo.sh <UPDATE_FNAME> <UPDATE_TYPE> <AUTO_SWAP> <SWAPPED_PARTID> <MAX_ITER> <STEP_SZ>

Output image results will be saved to classification/data_dict/shape_and_features/results/demo/images/.

Example 1

Automatically decide where to change, and change both shape and texture

cd classification/data_dict/shape_and_feature/
./scripts/edit_and_visualize_demo.sh 18.jpg shape_and_texture True 0 10 0.25

Example 2

Specify to change the top, and change both shape and texture

cd classification/data_dict/shape_and_feature/
./scripts/edit_and_visualize_demo.sh 3.jpg shape_and_texture False 0 10 0.05

Example 3

Specify to change the top, and change only the shape

cd classification/data_dict/shape_and_feature/
./scripts/edit_and_visualize_demo.sh 4.jpg shape_only False 0 10 0.25

License

FashionPlus is CC-BY-NC 4.0 licensed, as found in the LICENSE file.

Citation

If you find this useful for your research, please consider citing:

@inproceedings{hsiao2019fashionplus,
  title={Fashion++: Minimal Edits for Outfit Improvement},
  author={Hsiao, Wei-Lin and Katsman, Isay and Wu, Chao-Yuan and Parikh, Devi and Grauman, Kristen},
  booktitle={In Proceedings of the IEEE International Conference on Computer Vision (ICCV)},
  year={2019}
}

Acknowledgements

This code borrows heavily from pix2pixHD and BicycleGAN.