1. Explore ImageNet. ImageNet sample images, Kaggle ImageNet Mini 1000, 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?

   

                                                 *Picture of ball player from imageNet*
   

   I tried to download the dataset and found out it will take more than 1 day since it’s over 1TB!

   So I just took a look at the sample images on GitHub. Quickly browsing the pictures, I found out that more of the pictures are taken in daily life with a lot of noise, and that there are lots of objects in a single picture. For example, in “Ball Player”, the player is very small in the middle and there are as many people wearing white shirts as he is. So I’m curious whether this picture will decrease the accuracy of imageNet when it comes to not only ball players, but also categories like “playground”, “white shirt”, and “crowd” (if they have included them).

   Meanwhile, it shows the importance of a good dataset to a training model.

I think imageNet protects the privacy to a great extent. Firstly, it’s a model for non-commercial and educational/research purposes. Secondly, there is a reminder message about the privacy policy on the home page. However, there will be some issues as so many classification models are trained based on imageNet, which might be used for commercial purposes. Also, people may upload images found on the internet which may cause copyright issues.

1. 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...

   1292_1695264946.mp4

   Experiment Results:

   Group 1 - Works within 10 try:

   Water bottle, remote control, cellular telephone, loudspeaker, desk

   Group 2 - Works more than 10 tries:

   Mouse, chair

   Group 3 - Never recognized successfully:

   Cable, tissue box, book, pen

   Experiment Record

   Items	Accuracy
   Water bottle	65%
   Remote control	35%
   iPhone	35%
   Speaker	25%
   Mouse	6%
   Office Chair	5%
   Cable	0%
   Tissue Box	0%
   Book	0%
   Pen	0%

   Possible factors:

   1. Position: If the item is in the center, it’s easier to be recognized.
   2. Light: Unless it’s really dark, the light does not matter a lot. (reflection matters)
   3. Background: The cleaner the background, the easier the recognition.
   4. Time: When I hold the water bottle in front of the camera for 1 minute, it’s easier to recognize.
   5. Shape: A regular item in daily life could be recognized, but items in special shape cannot.