Tfjs: Update PoseNet readme to give guidance on training/fine tuning options.

[tafsiri]

We have a number of questions on this and could do some reasearch into what the options are for folks using the pre-trained model who want to adapt it in some way (even if the answer is that there isn't a way to fine tune it/retrain it).

[oveddan]

+1. It would be great to let people train their own PoseNet models. I get emails almost weekly about this.

[Reiss-Cashmore]

+1 I would be interested in training a posenet model using a game engine like unity. That way I can easily script particular poses and animations on humanoid models, from all different angles. Then render that using an engine like unity and push the output frames through posenet with annotations or metadata powered from the values within unity about the humanoid joint positions and rotations.

Seems like there is some very interesting research and experiments in this area quite recently.

• https://medium.com/@laanlabs/real-time-3d-car-pose-estimation-trained-on-synthetic-data-5fa4a2c16634
• https://bitbucket.org/Unity-Technologies/ml-imagesynthesis
• https://arxiv.org/pdf/1803.00455.pdf
• https://heartbeat.fritz.ai/synthetic-data-a-bridge-over-the-data-moat-29f392a52f27