Human Parsing and its conditioned GAN

---

Human Parsing and its conditioned GAN

This is more of a "toy" project where I apply human parsing, a special kind of semantic segmentation on a fashion dataset. Conditioned on the parsing results, I use cGAN to generate synthetic data and impose the generated image to output same parsing results as the original image. Fig. 1. From left to right: (1)the original image, (2) the synthetic image, and (3) the human parsing results overlayed on the original image.

This is quite interesting because for some images, I notice the synthetic data has better parsing results than the original image. Fig. 2. In this figure, the 1st column is the original image, the 2nd column its parsing results; the 3rd column the synthetic image conditioned on the parsing results, the 4th column the parsing results of the synthetic data.

The overview of the network is illustrated in the figure below:
(innocent.jpg)

Well, I am writing a paper out of this. So...no, not now.
But here are some more synthetic images to "appreciate". Some faces are really freaking me out. Maybe I should condition it with some facial keypoints... But I want to save the hustle. That's it folks. Hope you have enjoyed it.