I did some experiments that really highlight how heavily image recognition algorithms depend on probability - and how unprepared they are for surrealism.


My student trained some classifiers on ImageNet and the result was that all quadrupeds were predicted as dog.

We then switched to MS Coco and found that -- while the caption generation is often good for a laugh -- the object detection was not pretty good in most cases.

Not detecting sheep on trees maybe shows that the Deep Networks now have actually a good "common sense"

I wonder how it fare on the columbine harvester picture?



@deeds Ha, what a picture! As you predicted, Microsoft Azure (trained on Coco, I believe) reported it as "a crowd of people"

Sign in to participate in the conversation
Wandering Shop

The Wandering Shop is a Mastodon instance initially geared for the science fiction and fantasy community but open to anyone. We want our 'local' timeline to have the feel of a coffee shop at a good convention: tables full of friendly conversation on a wide variety of topics. We welcome everyone who wants to participate, so long as you're willing to abide by our code of conduct.