4 ms·

Yeah, two nearly identical versions (probably only resolution differences) of our logo resulted in 21% & 94% similarity scores

• Jack000 9y ago
  it's definitely not perfect. I think the embedding-vector approach of image search isn't scale invariant.
  • dsnuh 9y ago
    Forgive 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 ago
      Considering all the bad logos in the world that humans make, use and like - I wouldn't bet on it.