Episode 500 - 8 Years Of Progress In Cloth Simulations! 👕

Dear fellow scholars, this is Two Minute Papers episode number 500 with Dr. Zsolnai. And on this glorious day, we are going to simulate the kinematics of yarn and cloth on our computers. We will transition into today's paper in a moment, but for context, here is a wonderful work to show you what we were able to do in 2012 and we will see how far we have come since this previous work was about creating these highly detailed cloth geometries for digital characters. Here you see one of its coolest results where it shows how the simulated forces pull the entire piece of garment together. We start out with dreaming up a piece of cloth geometry and this simulator gradually transforms it into a real world version of that by subjecting it to real physical forces. This is a step that we call yarn level relaxation. A few years ago when I worked at Disney Research, I attended to the talk of the Oscar award winning researcher Steve Marschner, who presented this paper. And when I saw these results, shockwaves went through my body. It was one of my truly formative hold onto your papers moments that I'll never forget. Now note that to produce these results, one had to wait for hours and hours to compute all all these interactions. So this paper was published in 2012 and now nearly nine years have passed. So I wonder how far we have come since. Well, let's see. Together today, with this new technique, we can conjure up similar animations where the pieces of garments tighten. Beautiful. Now let's look under the hood of the simulator and. Well, well, well. Do you see what I see here? Red dots. So why did I get so excited about a couple red dots? Let's find out. Together. These red dots solve a fundamental problem when simulating the movement of these yarns. The issue is that in the mathematical description of this problem, there is a stiffness term that does not behave well when two of these points slide too close to each other. Interestingly, our simulated material gets infinitely stiff in these points. This is incorrect behavior and it makes the simulation unstable. Not good. So what do we do to alleviate this? Now we can use this new technique that detects these cases and addresses them by introducing these additional red nodes. These are used as a stand in until things stabilize. Look, we wait until these two points slide off of each other and now the distances are large enough so that the mathematical framework can regain its validity and compute the stiffness term correctly. And look, the red dot disappears and the simulation can continue without breaking. So if we go back to another piece of under the hood footage, we now understand why, why These red dots come and go. They come when two nodes get too close to each other and they disappear as they pass each other, keeping the simulation intact. And with this method, we can simulate this beautiful phenomenon when we throw a piece of garment on the sphere and all kinds of stretching and sliding takes place. Marvelous. So what else can this do? Oh boy. It can even simulate multiple cloth layers. Look at the pocket and the stitching patterns here. Beautiful. We can also put a neck tag on this shirt and start stretching and shirring it into oblivion. Pay special attention to the difference in how the shirt and the neck tag reacts to the same forces. We can also stack three tablecloths on top of each other and see how they would behave if we would not simulate friction. And now the same footage with friction, Much more realistic. And if we look under the hood, you see that the algorithm is doing a ton of work with these red nodes. Look. The table notes that they had to insert tens of thousands of these nodes to keep the simulation intact. So how long do we have to wait for a simulation like this? The 2012 paper took several hours. So what about this one? Well, this says that we need a few seconds per time step and typically several timesteps correspond to one frame. So where does this put us? Well, it puts us in the domain of not hours per every frame of animation here, but to minutes and sometimes even seconds per frame. And not only that, but this simulator is also more robust as it can deal with these unpleasant cases where these points get too close to each other. So I think this was a great testament to the amazing rate of progress in computer graphics research. What a time to be alive. This episode has been supported by weights and biases. In this post they show you how to use their tool to draw bounding boxes for object detection and even more importantly, how to debug them. 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, 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 wnba.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.