MobileNet Image Recognition

[Lucas Mortensen]

Hi everyone

I am a student at DTU (Danish Technical University), and I am currently studying Artificial Intelligence. I am, at the moment, working on a project whereby I compared the two models MobileNet V2 and MobileNet V3 Large, expecting V3 to be better.

We made a test, where a phone was placed in the middle with a full background of no signal video (randomized black/white pixelation), In our tests, through 500 frames, V2 guessed correctly 80% of the time, while the V3 model only guessed correctly 9% of the time!

Any ideas, hypothesis or comments relating to this would be greatly appreciated. Thank you.