Question about label of DeepFashion

[XianWu18]

Hello~It seems that the label number of DeepFashion dataset is 8, which is different from common human parsing models. Do you merge the labels？What is the class of each label？

[raven38]

Hi, would you please refer to Appendix B. Datasets. In a section of deepfashion, the authors describe "we first get its segmentation map, then re-organize the map into eight categories: hair, face, skin (including hands and legs),top-clothes, bottom-clothes, socks, shoes, and background."

The label number of common human parsing models is 20. The author map 20 labels into 8 labels.

There a re-implementation of this paper (https://github.com/raven38/GroupDNet/blob/master/deepfashion.py). This implementation converts semantic labels with 20 labels into 8 labels in the code.

[XianWu18]

I have got it! Thanks for your reply! @raven38

[Ha0Tang]

Why didn't you use 20 labels, but 8 labels?

[jessemelpolio]

Because in terms of person image generation, it is not that necessary to consider every category. We think the father classes are sufficient enough. For example, upper clothes may contain T-shirts or long shirts. However, their most significant difference is the layout difference rather than their styles or textures, which we wish the latent code to capture after training. Consequently, we just merge some classes that usually appear in the same areas into one class.