Explore ImageNet. What surprises you about this data set? What questions do you have? Thinking back to last week’s assignment, can you think of any ethical considerations around how this data was collected Are there privacy considerations with the data?
I think the only way for ImageNet to make improvements in its accuracy is to collect more data. So the problem of ethics and privacy is inevitable. I guess we should just be careful of what we are posting on social media.
Using the ml5.js examples above, try running image classification on a variety of images. Pick at least 10 objects in your room. How many of these does it recognize? What other aspects of the image affect the classification, including but not limited to position, scale, lighting, etc.
- I first tried with my AirPods case. ImageNet was wavering between “perfume/essence” and “computer mouse”. I guess this is due to the shape of the case being too common and indefinite. However, when I opened up the case, the shape is still not recognizable enough and was being identified as “water pot”.
2. Then I tried using the AirPods themselves. I don’t know if it was due to my shaky hand, the result flickered from “microphone”, “bubble”, “spotlight”, to “vacuum”.
3. Since I got “perfume” from my first try with the AirPods case, decided to try actual perfume. It got it right
4. I tried another brand of cologne, and ImageNet got it right again.
5. I was still skeptical of its accuracy in identifying colognes, so I tried again with a different brand. This time it was never recognized as a perfume bottle but as “spotlight”, “acoustic guitar”, and “microphone”. I wondered if that was affected by the background.
6. Of course I tried another brand of cologne after. I realized that the result is heavily influenced by the size of the object. The bootle does look like a joystick when it’s closer to the camera.
7. I believe ImageNet’s webcam image classification has a lower accuracy due to the inability to correctly recognizing the concept of perspective and the difference between apparent size and actual size.
8. iPhone was identified correctly only when the screen is on.
9. Band-Aid was identified correctly only if it’s closer to the camera. I guess that has something to do with the lighting.
10. I couldn’t see the result without my glasses, so the angle/position of the object influenced the result.
