We’re aiming to identify emotions with something you might call a multimodal non-contact sensor rig. (I’ll detail it another time when our naive approach has faced the burden of testing and optimization). And you would have thought that the difficult task would be identifying the emotions from images, videos, etc. across multiple ethnicities, cultures, personalities and contexts. This is indeed going to be difficult, but the surprise here was exactly how bewildering “choosing emotions” is. Wikipedia, as ever, comes to the rescue with a short intro on emotion classification, but that really just the door opener to spend the next 6 months reading atop an ivory tower. At least it strengthens the case that Paul Ekman is the grandaddy of all facial image emotion recognition. So we’d be confident of the entry point for 6 months of reading. But to discover existing models, labeled image sets, open software and more is proving rather tricky. I am emotionally drawn to doing something uber simple like SAM or PAD emotion reporting. The real question is how the balance getting quick working results versus getting the foundation of something amazing and actually useful in the wild.
Today, I only have the beginning of this story – but I aim to have the middle of the story well underway before September.