Computer vision's hidden trick
There is also a well-hidden trick in the field of AI for convincing people that computer vision surpasses humans. That is, there are a lot of tricky and uncommon things in their image library (such as ImageNet), or the types of certain things are divided into very small categories, and you are asked to name it. For example, some strange animal in the ocean, some unusual plant, or let you name a specific species of a dog, a flower or even a whale. Confounding human testers with a large number of these images makes the neural network seem stronger.
It might indeed be easier for a computer to know a thing’s name if there are too many varieties of it. It's like a search engine, no one can surpass the ability of a search engine to retrieve information. However, for things that are familiar to people, computers cannot accurately recognize them every time. The top-1 recognition rate of neural networks is very low, which can cost lives. And even if they are recognized, the computer only knows their names, but does not know their properties, what their components and materials are.
Besides, is naming a dog's specific breed really a measure of visual ability? This is just playing with words, and computers are really better than humans at this. Few people have seen dogs of all breeds, but they can see a dog of an unknown breed and know it is a dog with relative accuracy. Computers don't have this ability, and if there is a special dog that has never appeared in the picture library, it will be difficult for it to know that it is a dog. Humans can also recognize dogs in abstract works of art or decorations. Even a three-year-old can do this. However, this is almost impossible for computers.
