Interesting that it ignored people on billboards but recognised a wristwatch as a 'clock'.

Also that billboard at the start is a viral ad for a tv show.

It didn't always. See 0:47... seems it rightly doesn't recognize black-and-white-photographed or cropped people.

Tint it red and that's pretty much how the Terminator sees. I'm just saying is all.

This is awesome.
I wonder if the detection as it is shown in the video is actually real-time, or if some of the information was added later in post.

That is a legit question. Ignoring that they used a segment of another video, the question is how quickly it used that video segment to render this. The keyword in the paper (https://arxiv.org/abs/1506.01497) is "Towards Real Time", which I guess means it may have been delayed a bit from the original video, but I didn't read past the abstract, and that paper is just the bases for the video.

To be of any real value, say in a car to help it self-drive, is how fast it can do it, and read street signs.

I'm going to guess the reason it doesn't do billboards is it is trained, for whatever reasons to only look so high. Then again, they took another video, then sent that video through the processor, so why ignore them? Also, can it recognize "storefront" "billboard" etc... reading those and translating in real time probably would be a bit advanced for now, but still...

0:35 tagged baby in stroller as 74% person, obviously that should be about 20%.

badum, tish!

I think it's this system:
https://www.ted.com/talks/joseph_redmon_how_a_computer_learns_to_recognize_objects_instantly

I'm especially amused when it highlights the handbags/backpacks....I just assume this is the prototype for Bender's stealing algorithm...

Comment hidden because you are ignoring dag.(show it anyway)

I love this kind of stuff. It's amazing that GPUs are the secret to good computer vision. Graphics cards are not only good at generating 3d worlds, but taking the real 3d world and converting it back to something virtual Fantastic.

Clearly the next step for this system is to instantly determine each person's age, occupation and annual income, along with with a quick factoid or two to add flavour:

https://www.ubisoft.com/en-us/game/watch-dogs/

The Machine vs Samaritan, it's coming!!

It is good to see that SkyNet's target recognition software is coming along, despite the delays based on the schedule we were taught.

Also, I am 10 angry ferrets in a suit, so I wonder how I would show up on this scan.