Body Part Segmentation and Colorization using Generative Adversarial Networks
- 0 Collaborators
Body Part Segmentation can achieve better results using GANs compared to traditional Segmentation techniques. We will further optimize it using Intel's OpenVINO Toolkit. Another scope of this project is Colorization of segmented images to generate Human-like images and study the working of GAN's. ...learn more
Project status: Under Development
Intel Technologies
AI DevCloud / Xeon,
Intel Opt ML/DL Framework,
OpenVINO
Overview / Usage
SEGMENTATION
Body Part Segmentation is widely used in Fashion Industry. Now with the influence of AI in fashion industry, traditional methods can be neglected and the proposed solution can be used for real time Segmentation.
GAN (Generative Adversarial Networks) are specifically used to achieve this task as they are very good at generalization and are accurate even when augmented data not present in training is tested.
Segmentation can be used to study body variations, types and sizes of people from various regions. It can be used to give a user buying Fashion accessories online, a real time feel of the same.
COLORIZATION
Another scope of this project is colorization which is right now only for the purposes of research. Basically what we are trying to study is how good and how fast is GAN able to learn features. What different changes in loss functions and GAN architecture is needed to be made, in order to make our GAN learn minute details of Human Body Parts.
We can use these GAN's to generate various different clothes for every individual and also know how it would look on them.
Methodology / Approach
The base architecture for GAN's are conditional GAN's. Further we have used deep convolutions in both Generator and Discriminator so as to get the essence of even the minutest of features.
For Segmentation
Generator uses U-net like architecture :
encoder: C64-C128-C256-C512-C512-C512-C512-C512
decoder: CD512-CD512-CD512-C512-C256-C128-C64
Discriminator uses : C64-C128-C256-C512 deep constitutional architecture.
**For Colorization **
The Generator and Discriminator architectures are still under trials. I would attach the best result we have attained so far.
Technologies Used
Tensorflow
**Intel's OpenVino ToolKit **
Intel AI DevCloud