Mokaba, Kennedy, Nikos discussed on The TWIML AI Podcast
We also use one of them's called months three point six am and the other one is called. MPI INFORMATICS TO BE IMPOSES mission days. The Three D. poses in the wild data set is that that you mentioned it use a particular kind of sensors that sensor and the kind of sparsity of that sensor similar to what you see in a system or is it you know somehow more more dance I'm curious about? The. Relationship between the primary way you're comparing performances based on Kennedy's sparse representations I wonder if that. Leaves something to be desired or an opportunity in terms of the ability to to kind of capture what's happening in complex motions with these sparse data points on the body. It's a good question. So we use their commercial Mokaba soup called accents uses these inertial measurement units. Ten or twelve on the body, but we don't just use that We also use a camera tracking the people around and we used to D- measurements were about the the joint locations in in those images to. Precisely align the estimated three D. body posed with the images and so we have A. A technique that fits again are simple imperial body model to these multiple kinds of measurements, the new measurements and the two D. measurements we also have to solve for the camera translation and rotation and things like that. But putting this altogether gives reasonable. We don't call it ground truth. We call it reference data. We're a little careful about that. You know it's not the same level as a motion capture lab. But you get the benefit of being outside in in the wild with all the complex lighting. And things that go on. A. Natural scenarios. A trade off the holy grail is to have perfect ground truth where you know every single bit of the motion with no crazy sensors to get in the way outdoors in the wild. But. It doesn't exist. So mom and mentioned earlier in this process of kind of experimenting and identifying. Things it didn't work with. Particular emphasis on the discriminator Nikos I'm wondering from your perspective. You know if you can share another example of something net, maybe thought was gonNA work and didn't inhale you had to adjust the approach to accommodate it. The experiment station proceeded was was the. But. Yeah there are some a few points. So we tried..