new google dataset,

[dobkeratops]

https://ai.googleblog.com/2018/09/introducing-inclusive-images-competition.html

whats the license on these images (would they require visible attribution to copy.. they seem to provide a credits list). i think they've got bounding boxes, and a fair number of labels , but the label lists dont look exhaustive yet... do they have scope for further annotation

https://www.kaggle.com/c/inclusive-images-challenge/data

even if you can't copy them, i guess it's just another label list to try and add to the graph.. it should be possible to train the same net on multiple image sets

[bbernhard]

Very cool, many thanks for sharing!

Looks like, they also made some progress -the last time I checked, they only had a github repo, but now there seems to be a full blown site (https://storage.googleapis.com/openimages/web/index.html). Very impressive!

Every time one of the big ones releases a dataset, I am always asking myself: "Are we on the right track? Does this even make sense? Should we allow other licenses besides CC0?..."

Don't get me wrong, we made some really great progress this year - would have never imagined that - but compared to google with its ~9M images it feels like "nothing".

But what I keep always forgetting is: we can't win against Google (or any other big company) when it comes down to sheer size. No matter, how hard we try, it's still a community driven project where people contribute in their free time. And that's fine. I think they are many more ways we can position ourselves and "win":

• better tooling (better integration of third party tools, ML libraries, querability, ..)
• truly open source (not company driven)
• friendly community
• some niche labels (I guess no matter how big your dataset is, it's never complete)
• only public domain images
• ...

I think the "only public domain images" is probably the most controversal point on the list. I also thought more than once about introducing differently licensed images. But I am a bit afraid that it gets really complicated then, to both maintain (the software) and use the dataset. I think it also gets pretty complicated for users then ("Can I use xy licensed data in my closed source application?"). In general CC4 licensed content requires you to attribute the content creator. But what about, if I use a neural net in my application, that was trained on hundreds of images. Do I need to attribute every image creator now?

In order to avoid such legal nightmares, I would prefer the CC0 ("do the f*ck you want") license. But that's just my personal opinion and as always not written in stone ;)

Speaking of public images: In case you've missed it: #166 Do you have any objections against that? If not, then I would upload the first archive to archive.org in course of the next week.

[dobkeratops]

But what about, if I use a neural net in my application, that was trained on hundreds of images. Do I need to attribute every image creator now?

thats a great point especially considering that these nets can sometimes regenerate the images

we can focus on making the most of the images we have I guess (continuing to refine the labels with attributes ..). I know what you mean about the 'drop in the ocean' effect but the logic stands to keep independent efforts going. I agree the query/seach stuff thats grown around the labelling system is an interesting dimension here.