This is a big one -- we need to create way to save poses associated with a particular conjecture. For v1, we're just making this a possibility, but I think as we build this thing out, we're going to need to think through the way we do this.
This also includes saving/training the models we want to use for pose recognition.
This is a big one -- we need to create way to save poses associated with a particular conjecture. For v1, we're just making this a possibility, but I think as we build this thing out, we're going to need to think through the way we do this.
This also includes saving/training the models we want to use for pose recognition.
For v1, this may just look like teachable machine.