I found that there are a lot of categories in ImageNet gallery are about animals, and they are classified by very detailed species, such as “robin” and “jay” instead of “bird”. But how to make sure that all the sample images are labeled correctly? And if there are images go to the wrong category, how would they affect the accuracy of the whole model?
It’s an interesting idea that find emojis around you, but the classifier is not so accurate. I found it’s easier to identify the identical objects, like hands, since everybody’s hands looks quite the same. It’s not working so well when identifying items that might be manufactured in different appearance. It asked me to find a hat, and I wore a irregular shape hat that day. It’s a hat but the program didn’t recognized it. I also found that the position and background are also great factors to the successes rate. It’s easier to identify images in the center of the canvas and with a clean background.
It says that I found a laptop, but there is no laptop on the screen.
I tried all four examples. For the two examples which classify an image, it does a better work with the image contains a single object, like a cat, than the image of landscapes.
I also tried the classifier with webcam, but I forgot to took pictures of them. I think the problems are similar to the emoji scavenger hunt game. Background affects a lot on the accuracy. I first show a pencil to the camera with a lot of stuff in the background. The program fail to recognize the pencil. Then I hold a white paper behind the pencil and block the messy background, it recognized the pencil successfully.