[vicbrooker]

I've been on-and-off interested in facial expression for a while, and can't help but feel that 'the emotion tool' is kind of flawed.

Unless there's been a groundbreaking shift in the last 12 months or so, this software will be using the Facial Action Coding System to recognize muscle contractions which are then interpreted as 'emotions'. FACS is an amazing tool (can't recommend learning it enough if you're interested in this stuff), but it still relies interpreting the context of the expression to determine an emotion. It's a notation system for expressions, not an identification system for emotions.

There's an off chance that their algorithm has crowdsourced their data which can be even worse because the emotions we display and the ones we feel aren't always the same. For example, I may force a smile to cover fear. FACS training can take 100+ hours because a lot of heuristics for identifying expressions need to be unlearned.

It's a bit of worry when https://www.projectoxford.ai/demo/emotion#detection uses mostly posed photos to demonstrate their software. The third photo contains perhaps one smile that shows signs of being genuine, the rest seem to be posed (I'd need higher resolution to be sure). If this third photo was accurately reading their emotions (cf. expressions), then the results should not be that they are happy.

Replacing the emotions with names for expressions would make me feel a little more comfortable about using this: 'happy' with 'smile', 'angry' with 'glare' and so on. But then it seems a lot less useful so I see why they've described/marketed it the way they did.