AVA: A Finely Labeled Video Dataset for Human Action Understanding | Hacker News Reader

AVA: A Finely Labeled Video Dataset for Human Action Understanding

research.googleblog.com

44 pointshurrycane6 comments

From the download page:

> The AVA dataset contains 192 videos split into 154 training and 38 test videos. Each video has 15 minutes annotated in 3 second intervals, resulting in 300 annotated segments.

So basically this is a couple of CSV files annotating 192 videos, which are hosted on YouTube. ava_train_v1.0.csv is about 7 MB.

> basically this is a couple of..

I would prefer accuracy over complexity any day.

the most interesting thing i found was "We use movies as the source of AVA".

while the datasets will only grow, movies are not realistic - they are by design faked, acted, well lit etc. While that is probably the best thing to do with a starting set i am waiting for the CNN/RNN to start saying (much like the early black female standford researcher who was not identified as human face) that person is not walking - i know walking, it's just like John Cleese.

this is what makes this dataset poor. other datasets mentioned in the blog are based off youtube which is more realistic. movie based datasets have perfect lighting, center the subject are almost never useful (e.g. HMDB)

YouTube/Flickr/etc are far from ideal data sources. Do dogs drive cars? Flickr has tons of photos of dogs driving cars, eating ice cream, and doing tons of other rare-for-dogs things. Ultimately whatever the raw data source is what matters is how well is it curated, and that’s always going to be a highly labor intensive job that can be done well or poorly regardless of the source of the images being curated.

these are not random, raw youtube video datasets. they are hand curated dataset in specific classes (like using mechanical turk). youtube has really diverse videos with different lighting, and real world scenarios which makes it an excellent dataset. Movie clips look great but models trained on them are useless in real world.

Reply on news.ycombinator.com