How To Applying Generative Adversarial Networks To...

How To Applying Generative Adversarial Networks To...

Applying Generative Adversarial Networks to 2D–Animation

Introduction: Throughout its history, creating animated media has always been a balancing act, as adding more individually–drawn images, called frames, gives a better–looking product but requires more time and effort. Computer–generated animation has further exacerbated this issue by automatically creating frames between two key frames, i.e. frames that separate distinct actions or scenes. The issue here is that the output appears too perfect when done this way, which makes it less desirable than traditional hand–drawn animation. Providing an automatic method to generate frames with hand–drawn quality will greatly benefit multiple industries and academia.

One of the most compelling ... Show more content on Helpwriting.net ...This will consist of 10–frame animation clips with a textual description of the scene, and this will likely require work to be outsourced to Amazon Mechanical Turk (AMT).

On AMT, participants can be paid a small amount to watch the clip and provide short descriptions of what they observed. Animated clips can be found on websites such as YouTube, and I expect that having at least 6,000 total frames should provide a solid starting point. After checking the AMT responses and the downloaded clips, the data set can be split sixty/forty into its initial training and evaluation subsets.

While the data set is being assembled, I will begin designing the initial architecture for the

CNN. This will include a literature review of the most recent peer–reviewed papers in computer vision so that my model innovates on the current state of the art. I will implement this model in the programming language Python with the Caffe deep learning framework, which I used during my UCF REU. If my architecture is based on an approach that already exists, I will attempt to use a pretrained model to improve its learning rate. I also desire to strengthen my background in computer vision, so I will take related classes during that fall and/or commit time to self–study the topic further. When the initial data set and model are ready, I will allow the CNN to train until it stops noticeably

... Get more on HelpWriting.net ...