1

I'm not very good at recognizing faces, so I knew it would be taxing to watch Bombshell (2019), since all three lead actresses (Charlize Theron, Margot Robbie, and Nicole Kidman) look very similar to me. But I'm such a fan of Margot Robbie I made the effort.

Cutting to the chase, there was a point (46m) where I didn't know who I was looking at, so I paused the video, clipped this from a screenshot...

enter image description here

...and uploaded it to Reverse Image Search on duplichecker.com. Which unsurprisingly didn't find the "original". But it offered "Find similar images in different search engines".

The first few different search engines all gave worthless results ranging from Bing's We couldn't find any results to Lensio showing me 4 different blond[e?] actresses (none were in Bombshell). But yandex.com returned this...

enter image description here

If you don't know, the picture is actress Charlize Theron playing media personality Megyn Kelly - who at the exact point where I paused the movie is agonizing between staying at Fox or moving to NBC. Yandex has apparently seamlessly moved from less than 3% of a single video frame in a movie to the real-life "celebrity news" context being depicted.

If this isn't a wild fluke, is it evidence of huge variation in competency across different free AI-powered image recognition services?

FumbleFingers
  • 119
  • 4

1 Answers1

2

Any ML model's evaluation has to come from an independent test set which is supposed to be simple random samples from the target population, therefore only one instance is not evidence of huge variation in competency across different models, only an experienced fluke. To gather evidence, you need to prepare much more test samples perhaps beyond 10,000 or much more for a face recognition model. And in your case the result difference may not be due to the underlying ML model, but some reference database containing people's information.

cinch
  • 11,000
  • 3
  • 8
  • 17