Problems with Image Search
Categorical object recognition is great
If we search for ‘truck’ on Flickr, we get the following.
So far, so good.
Attribute-based object recognition is not so great
If we search for ‘brown truck’, we get
Not so good. Google Image Search does better:
although we get some trucks that are clearly not brown.
Now, if we go for ‘brown toy truck’,
or ‘vintage brown truck’,
the additional adjective returns more search results that are incorrect.
Logical combinations of attributes are terrible
If we try ‘brown or white truck’ (or similar terms like ‘brown truck or white truck’),
Google Image Search gives us trucks which are both brown and white.
Searching ‘non-yellow trucks’
returns pretty much all yellow trucks.
Its attribute-based classifier is clearly not taking logical propositions into account.