16% false positive is hilariously bad, especially given how rare actual handguns will be.

This headline shouldn't be tolerated here.

On the other hand missing completely it in 3 out of 30 tests (samples with a number of selected "optimizations") is a "terrible" 10% is awfully worse (in an alarm system).

>For the experiments, we selected 30 scenes from the previously used videos with the next requirements. Each scene is made up of at least 5 frames, filmed in a fixed scenario, i.e., in the same place, and the pistols are clearly visible to a human viewer. ... The model successfully detects the pistol in 27 scenes with an average time interval AATpI=0.2 seconds, which is good enough for an alarm system. The detector fails to detect pistols only in three scenes.

I am probably particularly picky but (from the paper):

>This section analyzes the performance of our best detection model on seven public videos, with low qualities, downloaded from youtube. Six of them are pieces of well known films from the 90s, James Bon: The World is Not Enough (video 1), 3 pieces from Pulp Fiction (video 2, 3 and 4), Impossible Mission: Rogue Nation (video 5) and Mister Bin (video 6). The seventh video is a long pistol threatening video (video 7).

... if you cannot spell Bond or Mister Bean, and you classify Mission Impossible: Rogue Nation as a 90's movie (calling it Impossible Mission) you don't look like having been particularly accurate when writing the article, if the same accurateness is in the actual numbers, there is a problem.

Anyway the "final" sample (30 scenes) is so small that it makes very little sense to derive anything from it.

What happens when it pushes the panic button when a firearm is present, but carried without threat, legally?

That sort of accuracy seems like about three orders of magnitude off from what would be useful in the field.

Unless you just basically want a drug-sniffing dog, for guns. That is, something to "scienceify" something that would otherwise be problematic (searching this car, or shooting that suspect).

"Hey judge, the DOG said it, it must be true!"

"Hey judge, the COMPUTER said it, it must be true!"

That number should be compared to the rate human officers unnecessarily escalate a situation because they think there's a gun. I don't know that number but it could easily be higher than 16%.

It should really be compared to the number of instances human officer's act as if there could be a gun " just in case ". Like in a traffic stop.

If officer's don't have to potentially gun someone down because they made a hectic movement 84% of the time, that's a pretty big win.

Imagine 84% of " black guy mysteriously died" actually led to convictions of officers because their ai thing told them "definitely no gun, bro"

> Imagine 84% of " black guy mysteriously died" actually led to convictions of officers because their ai thing told them "definitely no gun, bro"

I imagine that, like body cameras, the AI thing will, in those events, conveniently be deactivated, malfunctioning, misplaced, or have it's records lost.

probably. but you can only pull that stunt so many times.

> probably. but you can only pull that stunt so many times.

How many? Because apparently we haven't yet reached it with body cams.

Until you've produced a sample. Then some statistician can write a paper on it and someone takes it to the supreme court and 20 years later they will decide that there was probably something fishy going on.

Like I know?

But if you compound different systems, like the ai handgun thing and dont store the data in the same place, it becomes exponentially less likely that suddenly 2 critical systems failed at the same important moment for some reason.

the real issue with police brutality is the fact that it is, in fact, legal, and police is trained to respond in such a way. what protesters get wrong about police brutality is that its supposedly illegal. it is not. most instances of police killing someone are actually legal. which doesn't make it right. but its legal. saying somethings illegal when its not just makes you look like an idiot.

people need to protest for better laws, not getting individual cops sentenced, because that wont ever happen with any kind of regularity.

It could easily be lower than 16% too. Much lower.

Right, I'm saying I don't know the number, but this algorithm shouldn't be judged against a theoretical perfect officer, because that's not reality.

There are also a lot of legal gun owners who have a gun but are not a threat.

They seem to be spotting visible guns, not concealed ones, and the focus is on guns being brandished instead of holstered.

This is preliminary work but it wouldn't be useful to have an alarm go off every time another police officer or security guard comes into view with a gun on his hip.

That's true in the US, but not in many other places (the UK for example).

I guess you'd have to compare the failure rate to that of humans, similar to automated driving.

Beside that, the camera would have to view the subject from the correct angle, so, from the back, most likely position if they're carrying concealed.

The usefulness of a system like this would be where the police have a suspect cornered with their weapons drawn, but the suspect refuses to comply with instructions. If the suspect makes a suspicious move (e.g. as if he is reaching to pull a gun from a back pocket) the police will shoot immediately, since they would not have time to react if they waited to see if he was pulling a gun or not. Right now, the human system provides a 100% false positive rate. A weapon system that incorporates this detector would allow the police to pull the trigger when the suspect makes the suspicious move, but delay the firing of the weapon until the system detects a gun. Even with a 16% false positive rate it would save lives.

That said, I don't think police would want it if it might increase the risk to their lives.

Adding any kind of electronic firing control mechanism to a service pistol is a non-starter for a bunch of reasons:

MTBF: Even small guns produce recoil, which over time will damage or degrade any electronics built into the firing mechanism.

Security: if a "computer" of whatever complexity is controlling the firing of the gun it presents surface area both for attacks and bugs.

Rank and file acceptance: No officer will want to go into the field with a service weapon fitted with "safety restrictions" or "casualty reduction mechanisms" other than the trigger safety.

I agree with everything you said, except MTBF.

The training guns could be separate. Surely the "on-duty" guns are not being fired often enough that this will be an issue at all... right? Right??

#1 feeds into #3-- How confident are your rank and file going to be when the "training" guns fail regularly. Also there is something to be said for training with the exact same weapon you plan to use in the field.