Abstract
The project’s goal is to optimize the very heavy computationally and time speaking high resolution rendering process, through the rendering of low resolution images and the generation of high resolution images using super resolution techniques. The goal of the project is to accelerate the process of high resolution image generation using generative adversarial networks in a computational effective way
Challenges
- Understanding the startup dynamic as DataGen integrates different projects into the main product.
- Understand classical imagery techniques and their impact in the Computer Vision field.
- Work with a huge amount of data. Images and especially high resolution ones are very heavy and that really demands consideration for the use of algorithms.
- Deal with the resources limitation compare to the vision demands.
- The use of Generative Adversarial Network which is a very new techniques with still a lot to discover and research.
- The translate the human impression of high resolution into mathematical meaning for the model to be trained.
Achievements (according to KPIs)
- Working super resolution pipeline
- Successful transformation from low samples rendered images to bicubic low resolution images using Unet.
Further development
The major development should be the complete training of the GAN for super resolution that fits DataGen demands. As it is very demanding and not at all a precise technique , we worked by training GAN’s specific parts and that could be improved.
Also, the work was mostly done on natural movements of synthetics hands , this could be a complete new challenges to scale it to synthetics faces.
