Yes, the idea is that the model is local to the user and runs in their browser (pushing the computation to the edge). Here's a project demonstrating pose estima

Yes, the idea is that the model is local to the user and runs in their browser (pushing the computation to the edge). Here's a project demonstrating pose estimation from Twilio's engineering blog: https://www.twilio.com/blog/pose-detection-video-tensorflow-js https://www.twilio.com/blog/pose-detection-video-tensorflow-.... Re: bandwidth. Do you mean the weights? If so then yes, you'd have to think about distilling large models to smaller ones. There are techniques for doing this. Here is an explanation of distillation from Floyd Hub's engineering blog: https://blog.floydhub.com/knowledge-distillation/ https://blog.floydhub.com/knowledge-distillation/.