I just stumbled upon Flemish Scrollers (The Flemish Scrollers, 2021-2022 – Dries Depoorter), a fun tool that observes livestreams of the Flemish parliament in Belgium and notifies Twitter “Hey @[Politician Handle], put down the phone and pay attention” or some such with an image of the person doodeling instead of looking up bored as hell.
What I would like to know is whether anyone of you can find sources/has labeled data for something like this since I would actually be interested to something for my Parliament as well; I already checked and they seem to be live streaming their meetings as well, so this should be the smaller issue. Labeling everything would be the harder part.
Of course, data for inattentiveness on such a scale (room for 100+ people) should be rare to find; thus, if nothing in that order exists, other ideas that could be used as a basis.
- the Flemish Scroller account posts images/videos to twitter - good ground truth of inattentiveness
- Transfer Learning from the distracted driver dataset on Kaggle. I am not sure this will provide good results since that data is much closer to the person than a video of the Parliament. Anyone care to guess whether this is a viable and wortwhile approach?
- Few shot-learning with attentive people as good example, everyone who does not match the good example (ie. outliers) gets scolded in the future?
I would be happy about any input on this topic!