Their method is detailed in the whitepaper [1], mainly section 3.2. TL;DR mturk for high level annotations (frame contains/does not contain object in class), and internal annotators for bounding boxes.
210,000 annotated videos is amazingly large, and as others asked, I really wonder how they did their training - probably offloaded it to users more than likely. Really glad to see these data sets going out to the world.
It is a huge problem that Google/Apple/Facebook/Amazon have the bulk of the data gathering ability, which just means they have 100x the ability to iterate machine learning capabilities and applications. For them it's a huge win and exactly what I would be doing if I were them, but it leaves startups and smaller machine learning companies like us out there scrambling for new data sources that aren't touched by those (hint: there basically are none).
edit: This data also really helps us start teaching our machine vision systems mechanical and physical processes, as described here: https://arxiv.org/abs/1603.01312
It seems odd that only one object of each class is tracked in a video segment. I'd imagine that it would be a limitation for algorithms that generate bounding boxes for each class, as they might be penalized for correctly finding the non-tracked instances of a given class. Is only tracking a single instance of a class standard for this kind of dataset?
All this amazing automation and deep learning and Google still spams me with pewdiepie and other meme lord YT drama garbage suggestions despite clicking "not interested" hundreds of times.
8 comments
[ 4.5 ms ] story [ 42.9 ms ] thread[1] https://arxiv.org/pdf/1702.00824.pdf
It is a huge problem that Google/Apple/Facebook/Amazon have the bulk of the data gathering ability, which just means they have 100x the ability to iterate machine learning capabilities and applications. For them it's a huge win and exactly what I would be doing if I were them, but it leaves startups and smaller machine learning companies like us out there scrambling for new data sources that aren't touched by those (hint: there basically are none).
edit: This data also really helps us start teaching our machine vision systems mechanical and physical processes, as described here: https://arxiv.org/abs/1603.01312
One of them is a guy drinking what appears to be urine from a toilet with a spoon.