Image summary that cannot be recognized correctly even with the latest AI image recognition model
Image recognition technology using artificial intelligence (AI) is active in various fields such as self-driving cars and security systems, but it is an image recognition model '
What Inception Net Doesn't See – Abubakar Abid – Zou Group @Stanford
https://abidlabs.github.io/Inception-Blindspots/
◆ 1: Upside down car
If the car is upside down, the image recognition model will display the image as an 'analog clock' or ...
I misunderstand it as a 'tow truck'.
◆ 2: Sliced apple
InceptionNet can correctly recognize even cut fruits, but in the case of apples, 'cucumber' or ...
It is misidentified as 'banana'.
◆ 3: Pakistani groom costume
The costumes that the Pakistani groom cuts are 'fur coats' and ...
Misidentified as 'gas mask'.
◆ 4: Long balloons
InceptionNet does not seem to recognize balloons other than circles as balloons, and long and thin balloons that are likely to be used in balloon art are misidentified as 'hammers'.
In addition, the balloon art dog was mistaken for a 'maraca'.
◆ 5: Mask
The image dataset for training InceptionNet was collected before the pandemic of the new coronavirus, so it seems that the mask could not be recognized correctly. The three-dimensional mask is in a 'bucket' ...
The unused mask was mistaken for a 'purse'. Did you mistake the string on your ear for the handle of the handbag?
◆ 6: Illustrated lion
InceptionNet also seems to be not good at recognizing illustration images. In particular, it seems that he is not good at correctly recognizing the lion in the illustration, and the following illustrations are 'bloodhound' and ...
It seems that it has been judged as an 'envelope'.
◆ 7: Roomba
In addition, InceptionNet seems to recognize the cleaning robot Roomba as a 'CD player'. Certainly, if you were shown this shape before the birth of Roomba, you might mistake it for a CD player.
◆ 8: Flying goose
The appearance of geese flying in a flock is mistakenly recognized as a 'war plane'.
'InceptionNet is considered to be very accurate, and we have released an API using InceptionNet that can be used by multiple companies for the general public, but there are problems with InceptionNet,' said Abid. Many people don't understand that. '
Even in the experiments actually conducted by Mr. Abido, 'objects in abnormal positions (◆ 1)', 'non-Western items (◆ 3)', 'non-standard artistic images (◆ 6)', etc. It turns out that it is easy to be misrecognized. Avid argues that the image recognition model needs to be thoroughly investigated and documented for multiple such issues and made known to users for their understanding.
Related Posts:
in Software, Posted by logu_ii