Week 2 Assignment
Exploring ImageNet
When I looked at ImageNet, I was surprised by how many different results came up after looking up a single word. For example, when I entered “tiger” in the search engine, it not only showed images of the cat tiger but also tiger shark and tiger beetle. This indicated that you have to be specific in what you search for in ImageNet if you are looking for a specific image.
In terms of privacy, I wonder whether the image source knows their photos/images are being used. Also, I feel like it would become a problem if a person shows up in an image and the image is labelled as something unfavorable.
Testing out image classification on ml5.js
I tried the image classification using several examples. I tried the file upload and the camera (video) methods and discovered that uploading image files leads to better recognitions than placing the item in front of the camera. When I uploaded a picture of myself, the ml5 recognized me as a wig, which I considered pretty close considering that it only detected my hair. But when I put myself in front of the camera on a white background, it recognized me as a lab coat, which obviously isn’t even close to correct. I even put my face really close to the camera.
Other than my face, I tried out the following items. The results are next to them:
- Perfume — perfume, essence 0.9
- Stuffed animal — teddy bear, 0.4
- Tissue box — toilet tissue, bathroom tissue, 0.4
- Cup — Beer glass, 0.3
- Headphones — microphone, mike, 0.4–0.5
- Pen —ballpoint pen, 0.2–0.5
- Lotion — lotion, 0.7
- Eye Drops — syringe, 0.1–0.3
- Bag —purse 0.5
- T-Shirt — lab coat, 0.1–0.4
It seems that it recognizes mostly everything except my face. In order for it to recognize correctly, I realized that the item must be placed in the right angle, the lighting should not be too bright or dark, and there shouldn’t be anything else in the background (otherwise it would detect the other items in the background). When I held the pen vertically to the camera, it first recognized it as a pencil eraser, but when I held it horizontally, it recognized it as a ballpoint pen. I am assuming that whatever database the program is using, it mostly has images of pens that are in a horizontal position.
