Hi, I am curious to see how movenet performs for classification of yoga pose . I was thinking of using teachable machine to classify the poses . However if i understand correctly it only supports posenet. So i am wondering if there is already a model that classifies different poses and try to run a webapp with p5js . I tried to look into tensorflow model hub but could not find any particular models as such
#special-interest-groups:tfjs #movenet
Hi there @tim_tensor (great name by the way)
So as you saw Teachable Machine while a wonderful educational tool is getting a bit old in terms of state of the art model usage. We do not recommend using PoseNet for this, instead something like MoveNet or MediaPipe’s Blazepose model would be a better bet. However both these models give you they key points but not the classification of key points for a given pose.
So next question is how do you build on that?
Well to save repeating myself I actually cover how to make your own teachable machine here:
While this is for image recognition task, the principle TM uses behind the scenes for pose is similar. Instead of using MobileNet as the base model, you would use one of the above 2 that output the 17 keypoints, and instead you would feed those keypoints into the multi layer perceptron and train that on data from captured targets in different positions etc that represent the pose you want to recognize.
However to understand that video above you need to start earlier on in the course to understand the concepts of Transfer Learning, and Tensors and all that fun stuff - so if you are not confident in those areas, I recommend taking my course on that YouTube channel and by the time you finish chapter 5 you will have the knowledge you need to modify that to apply to your pose classification task above!
Hope that helps!