Virtual Bones Make Everything Better! 💪

Dear fellow scholars, this is Two Minute Papers with Dr. Zsolnai. Today we are going to see that virtual bones make everything better. This new paper is about setting up bones and joints for our virtual characters to be able to compute deformations. Deformations are at the heart of computer animation. Look, all of these sequences require a carefully designed technique that can move these joints around and simulate its effect on the entirety of the body. Things move around, stretch and bulge. But there is a problem. What's the problem? Well, even with state of the art deformation techniques, sometimes this happens. Did you catch it? There is the problem. Look. The hip region unfortunately bulges inwards. Is this specific to this technique? No, no, not in the slightest. Pretty much all of the previous techniques showcase that to some effect. This is perhaps our most intense case of this inward bulging. So let's have a taste of the new method. How does it deal with a case like this? Perfectly, that's how. Loving it. Now hold onto your papers, because it works by creating something that the authors refer to as virtual bones. Let's look under the hood and locate them. There they are. These red dots showcase these virtual bones. We can set them up as a parameter and the algorithm distributes them automatically. Here we have 100 of them, but we can go to 200. Or if we so desire, we can request even 500 of them. So what difference does this make with a hundred virtual bones? Let's see. Yes, here you see that the cooler colors, like blue, showcase the regions that are deforming accurately. And the warmer colors, for instance, red, showcase the problematic regions where the technique did not perform well. The red part means that these deformations can be off by about 2cm, or about 3/4 of an inch. I would say that is great news, because even with only a hundred virtual bones, we get an acceptable animation. However, the technique is still somewhat inaccurate around the knee and the hips. If you are one of our really precise fellow scholars and feel that even that tiny mismatch is too much, we can raise the number of virtual bones to 500. And let's see. There we go. Still some imperfections around the knees, but the rest is accurate to a small fraction of an inch. Excellent. The hips and knees seem to be a common theme. Look, they show up in this example too. And as in the previous case, even the 100 virtual bone animation is acceptable. And most of the problems can be remedied by adding 500 of them. Still some issues around the elbows. So far we have looked at the new solution and marked the good and bad regions with heatmaps so now how about looking at the reference footage and the new technique side by side? Why? Because we'll find out whether it is just good at fooling the human eye or does it really match up. Let's have a look together. This is linear blend skinning, a state of the art method. For now, we can accept this as the reference. Note that setting this up is expensive both in terms of computation and it also requires a skilled artist to place these helper joints correctly. This looks great. So how does the new method with the virtual bones look under the hood? These correspond to those. So why do all this? Because the new method can be computed much, much cheaper. So let's see what the results look like. Hmm. Yeah. Very close to the reference results. Absolutely amazing. Now let's run a torture test that would make any computer graphics researcher blush. Oh my. There are so many characters here animated at the same time. So how long do we have to wait for these accurate simulations? Minutes to hours? Let me know your guess in the comments below. I'll wait. Thank you. Now hold on to your papers because all this takes about 5 milliseconds per frame. 5 milliseconds? This seems well over a hundred characters rocking out. And the new technique doesn't even break a sweat. So I hope that with this computer animations are going to become a lot more realistic in the near future. What a time to be alive. Also, make sure to have a look at the paper in the video description. I loved the beautiful mathematics and the clarity in there. It clearly states the contributions in a bulleted list which is a more and more common occurrence. That's good. But Locke even provides an image of these contributions right there, making it even clearer to the reader. Generally, details like this show that the authors went out of their way and spent a great deal of extra time writing a crystal clear paper. It takes much, much more time than many may imagine. So I would like to send a big thank you to the authors for that. Way to go. This episode has been supported by weights and biases. In this post they show you how to use transformers from the Hugging Face library and how to track your model performance. During my PhD studies I trained a ton of neural networks which were used in our experiments. However, over time there was just too much data in our repositories and what I am looking for is not data and but insight. And that's exactly how weights and biases helps you by organizing your experiments. It is used by more than 200 companies and research institutions including OpenAI, Toyota Research, GitHub, and more. And get this Weights and Biases is free for all individuals, academics and open source projects. Make sure to Visit them through wnb.com papers or just click the link in the video description and you can get a free demo today. Our thanks to WeightsAndBiases for their long standing support and for helping us make better videos for you. Thanks for watching and for your generous support and I'll see you next time.