Exercises from A People’s Guide to AI.

Try running image classification on a variety of images.

I tried them with some photos I took. I noticed the model tends to prefer images that have a distinct, ordinary object and no humans. The quality of images also matters.

unique building cannot be recognized

image.png

a standard traffic light

image.png

There are some good images that the model failed to recognize. I guess it is because of the dataset. Maybe they didn’t include those objects.

image.png

image.png

I’m surprised that it recognize this picture, because it is blurry and the train is hidden behind the light.

image.png

I took these two pictures almost at the same time, while the exposure is different. Comparing the results, I think the lightness and exposure of images matter a lot.

image.png

I’m amazed by its ability to recognize different kinds of dogs. Over 90% of the dog pictures I tried with get the correct result.

image.png

image.png