[flatline]

I also noticed some false positives. It is a good demonstration of the difficulty of the task: there is no easy, automated, out of the box solution that will arrive at the correct answer. You could make a decent approximation if you could correlate frames and track individual bottles but even that is beyond the scope of this demo.

[Fragoel2]

Not to mention the 90% recall, meaning that the model is missing 10% of the bottles in the annotated data.

Cool blog post, cool tutorial, and great tool but its not going to win any contest where you need to guess an exact number.