Thoughts on ImageNet
While for some reason I wasn't able to see the images themselves (the page was stuck perpetually loading them), I did look through many of the categories and was quite interested by the priorities of the data set. It seemed to have an unexpected preference towards the natural world at least within the heirarchial category structure, with Flora/Fauna/Geoology/Fungus/Natural Object making up the majority of the top level folders. I'm not sure how many ethical and privacy considerations can be raised with the usage of commercial photos of objects/nature, as while they don't have rights to the images, the photos themselves are fairly innocuous and not used/seen as individual pictures.
Experiments with MobileNet
For my experiments with MobileNet and the ML5 pre-trained classification models, I decided to make use of a variable color temperature and intensity light at my desk to test the algorithm's preferences towards lighting. For control, I first tested the model's ability to detect objects with the room lights on full. MobileNet was able to detect all objects except for a stapler and a fork (which it thought was a spatula). I then turned off the room lights and turned on the desk light, which sat directly behind the camera so as to control shadow direction as much as possible. Each object was put in four lighting conditions: warm and bright, warm and dim, cold and bright, and cold and dim.
With an increase in intensity nearly always came an increase in confidence score (typically along the order of about 0.05-0.1 points). More interestingly however, the color temnperature results changed depending on the object. While the water bottle got on average 0.15 point confidence boost from cool white lighting, the banana experienced on average a 0.25 point detrement.