4 ms·
Cool! I tried it with our logo (https://www.airbornos.com/images/logo-mark.png https://www.airbornos.com/images/logo-mark.png), which is super generic (I spent
by twiss 9y ago
Cool! I tried it with our logo (https://www.airbornos.com/images/logo-mark.png https://www.airbornos.com/images/logo-mark.png), which is super generic (I spent $0 on it). It came up as 96% unique, even though there's two icons of clouds in the list of similar icons. Furthermore, it would be cool if it looked for similarity with a database of other logos as well, and not just icons.
Other small point: maybe you could show the uploaded logo again next to the other logos in the "Color/Contrast" and "Overall" categories, to (probably) drive home the point how bad your logo is in comparison.
- dschep 9y agoYeah, two nearly identical versions (probably only resolution differences) of our logo resulted in 21% & 94% similarity scores
- Jack000 9y agoit's definitely not perfect. I think the embedding-vector approach of image search isn't scale invariant.
- dsnuh 9y agoForgive me if this is a naive question, as I am not well versed in ML. Would it be possible to train using Mechanical Turk so that you get real human feedback and perception of quality?
- wingerlang 9y agoConsidering all the bad logos in the world that humans make, use and like - I wouldn't bet on it.