Dress Code Virtual Try-On Dataset

This repository presents the virtual try-on dataset proposed in:

D. Morelli, M. Fincato, M. Cornia, F. Landi, F. Cesari, R. Cucchiara
Dress Code: High-Resolution Multi-Category Virtual Try-On
Under Review

[Paper] [Dataset Request Form]

By making any use of the Dress Code Dataset, you accept and agree to comply with the terms and conditions reported here.

Dataset

We collected a new dataset for image-based virtual try-on composed of image pairs coming from different catalogs of YOOX NET-A-PORTER Group.
The dataset contains more than 50k high resolution model clothing images pairs divided into three different categories (i.e. dresses, upper-body clothes, lower-body clothes).

Summary

• 53792 garments
• 107584 images
• 3 categories
  • upper body
  • lower body
  • dresses
• 1024 x 768 image resolution
• additional infos
  • keypoints
  • label_maps
  • skeletons
  • DensePose

Additional Info

Along with model and garment image pair, we provide also the keypoints, skeleton, image label map, and densePose.

More info

Human Joints

For all image pairs of the dataset, we stored the joint coordinates of human poses. In particular, we used OpenPose [1] to extract 18 keypoints for each human body.

For each image, we provided a json file containing a dictionary with the keypoints key. The value of this key is a list of 18 elements, representing the joints of the human body. Each element is a list of 4 values, where the first two indicate the coordinates on the x and y axis respectively.

Human Skeletons

Skeletons are RGB images obtained connecting keypoints with lines.

Human Label Map

We employed a human parser to assign each pixel of the image to a specific category thus obtaining a segmentation mask for each target model. Specifically, we used the SCHP model [2] trained on the ATR dataset, a large single person human parsing dataset focused on fashion images with 18 classes.

Obtained images are composed of 1 channel filled with the category label value. Categories are mapped as follows:

 0    background
 1    hat
 2    hair
 3    sunglasses
 4    upper_clothes
 5    skirt
 6    pants
 7    dress
 8    belt
 9    left_shoe
10    right_shoe
11    head
12    left_leg
13    right_leg
14    left_arm
15    right_arm
16    bag
17    scarf

DensePose

We also extracted dense label and UV mapping from all the model images using DensePose [3].

Experimental Results

Low Resolution 256 x 192

Name	SSIM	FID	KID
CP-VTON [4]	0.803	35.16	2.245
CP-VTON+ [5]	0.902	25.19	1.586
CP-VTON' [4]	0.874	18.99	1.117
PFAFN [6]	0.902	14.38	0.743
VITON-GT [7]	0.899	13.80	0.711
WUTON [8]	0.902	13.28	0.771
ACGPN [9]	0.868	13.79	0.818
OURS	0.906	11.40	0.570

References

[1] Cao, et al. "OpenPose: Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields." IEEE TPAMI, 2019.

[2] Li, et al. "Self-Correction for Human Parsing." arXiv, 2019.

[3] Güler, et al. "Densepose: Dense human pose estimation in the wild." CVPR, 2018.

[4] Simonyan and Zisserman. "Very deep convolutional networks for large-scale image recognition." ICLR, 2015.

[5] Minar, et al. "CP-VTON+: Clothing Shape and Texture Preserving Image-Based Virtual Try-On." CVPR Workshops, 2020.

[6] Ge, et al. "Parser-Free Virtual Try-On via Distilling Appearance Flows." CVPR, 2021.

[7] Fincato, et al. "VITON-GT: An Image-based Virtual Try-On Model with Geometric Transformations." ICPR, 2020.

[8] Issenhuth, el al. "Do Not Mask What You Do Not Need to Mask: a Parser-Free Virtual Try-On." ECCV, 2020.

[9] Yang, et al. "Towards Photo-Realistic Virtual Try-On by Adaptively Generating-Preserving Image Content." CVPR, 2020.

Contact

If you have any general doubt about our dataset, please use the public issues section on this github repo. Alternatively, drop us an e-mail at davide.morelli [at] unimore.it or marcella.cornia [at] unimore.it.