Hacker News Reader

Well if you've already annotated all the bottles in the video then why do you need to train a model to do it?

He only hand annotated bottles in about 10% of the frames of the video. Those were used to train the model, which could then be used to annotate the rest of the frames.

to catch the ones that were invariably missed due to optical illusions etc.

wouldn't it also count reflections, for example, which are not real bottles to be counted ?

That's where I think they are going to use their discretion to "pick" the winner, despite somebody else having been just as correct.

For instance, they say the one he is holding in one scene isn't counted again in a later scene.

What about the reflection in his sunglasses? Is that 1 or 2 (or 3 if you count the source of the reflection wherever that is).

Rules specifically say reflections are counted, even calling out an example of reflections in sunglasses would count as additional bottles.

And un-count the ones that appear in more than one scene, as mentioned in the article

Reply on news.ycombinator.com