ImageNet: Analysis

The most surprising thing about the ImageNet dataset was probably the sheer volume of it. I was also very curious about the choice of words used to describe each dataset, particularly since the images are human-annotated. There are definitely ethical considerations when it comes to the fact that ImageNet does not own any of the images. Even though those images are free to view on the web anyways, the owners of the images have the right to know in what ways are their images being used and if they’re being used for machine learning algorithms.

Image Classification:

Plant → Shower Curtain

Glass → Water Jug

Plate → Cowboy Hat

Napkin → Handkerchief

Photo Frame → Desktop Computer

Highlighter → Lighter

Playing around with MobileNet felt frustrating as it did a poor job of recognizing any of the objects. I had to fix the angles and orientations several times before getting an answer similar to the object. As I could tell MobileNet is very dependent on the circumstances surrounding the object in the photo. I tried my best to clear everything else and it made a difference. The more the object took space from the background the more likely it was to be recognized. Another interesting observation I had was that there is a very narrow scope of pictures from other cultures. I added a picture of the Taj Mahal and it wasn’t able to recognize it even though it’s a quite infamous building. I then added a picture of the pyramids and it said it is a wreck. I reached the conclusion that MobileNet probably only contains generic images.