For the talk today, I will introduce the concept of model based coding. In particular, I will be discussing facial analysis and the concept of dynamic bandwidths. I will spend several slides on this material as my project has less to do with the particulars of algorithm as it does with the appropriate application of the algorithm. Next I will discuss the changes made to NSGA-II in order to make it more real time fro the application. In particular, I did this by combining it with a deterministic search. I will then present the results and conclude the presentation.
Mention the video games on essential parameter bullet Alternative to sending raw video footage Creation of “essential” parameters needed to reconstruct a scene Imagine the most recent multiplayer video games. The characters interact with each other, and the surroundings seemlessly. This is because the elements are based upon a model, not a video. A real-time analysis nightmare This is where the similarity ends, however. Video games take no real analysis. It is all built into the game software. For model based analysis, we need to find out what the user is doing by video processing instead of controller inputs.
Before discussing facial analysis it is good to get an idea of the power of model based coding. Take this scene for example, the participants in the conference are being transmitted extremely well. If you have ever participated in a conference on video, you know how bad it is. Other application include Gaming, where the character on the screen mimics your face, Man Machine interaction, where a computer can get an idea of you emotions using facial expressions. And video telephony. Most model based coding is at or below 1kb/s, and the last time I checked cellular communication, it was close to 9600 b/s at transmission and 1200 b/s when idle.
Okay here is how facial analysis by optimization is done. We need key animation parameters to manipulate the model of the face shown. The model based coder… But the analysis can be completed using synthesis…
The power of the coding is well illustrated from these example models… Ideally, I would have models as good as the ones seen here, but…
I have no funding for an expensive rendering software, and am a fairly dismal programmer. So this rendering will have to do.
The current research uses stochastic training to know when to do adjust based upon the error in the image. Gradient methods can be done in real time but catches up quickly with large movements of the head. To combat this, direct methods are used thast are not dependent on training. Direct methods are farther from being real time, but a more robust. TO make them more real time, researchers throw out FAPs. I agree that not all FAPs are needed fro realism, but currently only about 2 frames a second can be done using 12 parameters. FAP reduction is daunting, which FAPs are most important? For every facial scenario? And, no one is approaching this from a multiple objective standpoint.
What do I mean by dynamic bandwidth
I chose to use PSNR for its computational convenience. It is fast and easy to implement.
Use NSGA-II for the multiple objective optimization Assign a premature stopping criteria Choose bandwidth Select FAP sets Use deterministic algorithm
Have a two dimensional picture here, and explain the discrete line search