/DeepFashion_Try_On

Official code for "Towards Photo-Realistic Virtual Try-On by Adaptively Generating↔Preserving Image Content",CVPR‘20 https://arxiv.org/abs/2003.05863

Primary LanguagePython

Towards Photo-Realistic Virtual Try-On by Adaptively Generating↔Preserving Image Content, CVPR'20.

Official code for CVPR 2020 paper 'Towards Photo-Realistic Virtual Try-On by Adaptively Generating↔Preserving Image Content'. We rearrange the VITON dataset for easy access.

[Dataset Partition Label] [Sample Try-on Video] [Checkpoints]

[Dataset_Test] [Dataset_Train]

[Paper]

Inference

python test.py

Dataset Partition We present a criterion to introduce the difficulty of try-on for a certain reference image.

The specific key points we choose to evaluate the try-on difficulty

image

We use the pose map to calculate the difficulty level of try-on. The key motivation behind this is the more complex the occlusions and layouts are in the clothing area, the harder it will be. And the formula is given,

The formula to compute the difficulty of try-onreference image

image

where t is a certain key point, Mp' is the set of key point we take into consideration, and N is the size of the set.

Segmentation Label

0 -> Background
1 -> Hair
4 -> Upclothes
5 -> Left-shoe 
6 -> Right-shoe
8 -> Pants
9 -> Left_leg
10 -> Right_leg
11 -> Left_arm
12 -> Face
13 -> Right_arm

Sample images from different difficulty level

image

Sample Try-on Results

image

License

The use of this software is RESTRICTED to non-commercial research and educational purposes.

Citation

If you use our code or models in your research, please cite with:

@inproceedings{HanYang2020,
  title={Towards Photo-Realistic Virtual Try-On by Adaptively Generating↔Preserving Image Content},
  author={Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo and Ping Luo},
  booktitle={CVPR},
  year={2020}
}

Dataset

VITON Dataset This dataset is presented in VITON, containing 19,000 image pairs, each of which includes a front-view woman image and a top clothing image. After removing the invalid image pairs, it yields 16,253 pairs, further splitting into a training set of 14,221 paris and a testing set of 2,032 pairs.