Skip Navigation

Programmer tries to explain binary search to the police

365

You're viewing part of a thread.

Show Context
365 comments
  • Maybe I have no understanding of what a binary search is. My understanding is that you check halfway through the video, see if the thing has happened yet, then skip halfway to the end if it hasn't. Check again, skip again. When you see the cue that the event has happened, you rewind to halfway between the latest point where the event hadn't happened yet and the earliest point when it has. Keep doing that and you can pinpoint the exact frame where the event happens in a matter of minutes.

    Binary search would be largely useless in cases where you have a good chance of skipping right past the event. If the video is an hour long, and the event happens 34 minutes in and leaves a visual cue that lasts less than 11 minutes, then binary search does not find the event. At that point, watching the video fast forwarded would be the way to go, and that's not a binary search, that's just watching the video.

    So I should correct myself: the visual cue doesn't have to last the remainder of the video, it just needs to last until one of the points that you check. Which still makes it not useful for things that don't leave visual cues that last more than a few minutes, because it cannot find most of those events if they happen at a random time in an hour+ video.

    • When you see the cue that the event has happened, you rewind

      The event has happened, or the aftereffects that the event happened. That is my point, the aftereffects matter as much as the event itself. As long as the 'after' looks different than the 'before' for any reason, that is a marker to give you an indication on which way to go, rewind, or advance.

      And yes, either the effect or the aftereffects has to last long enough to be noticed by humans, less long by AI (faster to detect changes than humans). But the vast majority of events, when humans are involved, leave long aftereffects usually. Yes, not 100% of the time, but usually.

      • The event has happened, or the aftereffects that the event happened.

        In which case there are visual cues and it's something that the comment you argued with acknowledged would be eligible for binary search

        But the vast majority of events, when humans are involved, leave long aftereffects usually. Yes, not 100% of the time, but usually.

        Nobody said otherwise, you're arguing with strawmen

        • But the vast majority of events, when humans are involved, leave long aftereffects usually. Yes, not 100% of the time, but usually.

          Nobody said otherwise, you’re arguing with strawmen'

          Yes, they have. They've used it as a reason why a binary search would not work, that the event duration would be too short to be detectable.

          And that's not a strawman, that's making my point, that its not just the event, but the aftereffects of the event, that makes a binary search possible.

      • less long by AI (faster to detect changes than humans).

        Many things change things. A bit of smoke in the air might have been from a gunshot that happened 10 minutes ago, or it might have been from a cigarette 15 minutes ago. Binary search relies on changes that indicate a specific thing has happened–a broken window, a bike no longer there, blood stains on the street. Anything undetectable by humans would still be useless to AIs. A bit a smoke? Could have been a gunshot 3 minutes ago, could have been a cigarette, could be fog, could be a vape. Even the things that AIs are truly useful for, like interpreting video compression artifacts, wouldn't help, because any number of things can cause compression artifacts. How could it tell what pixels are slightly off color because of a gunshot 3 minutes ago, and what pixels are slightly off color because someone walked past the camera?

        At that point, just feed the entire video to the AI and have it tell you when it sees guns or puffs of smoke or hears screams. Binary search is useless when you can just have a machine watch the entire video in one sitting over the course of five seconds and tell you when the interesting thing happens.

        • Anything undetectable by humans would still be useless to AIs. A bit a smoke? Could have been a gunshot 3 minutes ago, could have been a cigarette, could be fog, could be a vape.

          Actually, an AI could determine the difference between those, based on shape, location, and opacity, etc.

          At that point, just feed the entire video to the AI and have it tell you when it sees guns or puffs of smoke or hears screams.

          Is there a point where one technique works better than another technique? Sure. I'm not arguing that. But if you're dealing with a very long time, you'd still want to do a binary search first.

          Binary search is useless when you can just have a machine watch the entire video in one sitting over the course of five seconds and tell you when the interesting thing happens.

          Depends on how long that tape is, which is what was being originally discussed by the OP.

          A binary search assisted by AI in determining the point in the tape where the effect happened quickly is still a very fast way of doing so (assuming the tape duration is very long), as alluded by others in other topic trees in this topic.

          • Actually, an AI could determine the difference between those, based on shape, location, and opacity, etc.

            Lmao now I know you're fucking with me

            Yeah lemme spend three weeks training this AI on the difference between gunsmoke, cigarette smoke, vapes, and fog in this specific alley. Oh, y'all already found the killer because someone just watched the video? Well my point stands, the AI could do it faster

            Once it's trained

            In another week

            Oh shit, it thought that guy's cell phone was a gun. See you in another month!

            • Actually, an AI could determine the difference between those, based on shape, location, and opacity, etc.

              Lmao now I know you’re fucking with me

              Yeah lemme spend three weeks training this AI on the difference between gunsmoke, cigarette smoke, vapes, and fog in this specific alley. Oh, y’all already found the killer because someone just watched the video? Well my point stands, the AI could do it faster

              Once it’s trained

              In another week

              Oh shit, it thought that guy’s cell phone was a gun. See you in another month!

              Um, I was being completely serious. Having AI determine shapes/opaqueness is a simple matter for it. And I'm assuming the training would already be done before the event happens, over time.

              You don't think crime forensics labs won't be training AI to do these kind of detections going forward? Really?

              (Maybe its a matter of people not truly grocking what AI will do and how it will change things, going forward. /shrug)

              • Having an AI search for shapes an opaqueness is still totally useless for a binary search if those semi-opaque shapes happen for 10 minutes 34 minutes into an hour long video

                Again, you'd just feed the whole video to an AI, you wouldn't have it do a binary search

                • Having an AI search for shapes an opaqueness is still totally useless for a binary search if those semi-opaque shapes happen for 10 minutes 34 minutes into an hour long video

                  Well one of those shapes would happen at the time of the event though, so it's not useless. One of those would be a gunshot smoke, and could be flagged for review.

                  Again, you’d just feed the whole video to an AI, you wouldn’t have it do a binary search

                  One day, when computers and AI are powerful enough, this will be the answer, but even then I would like to think behind the scenes they would use a binary search to speed up the processing time.

                  • The time of the event doesn't necessarily coincide with any of the times that you're checking. That's the whole point of looking for visual cues. Again, if the event happens 34 minutes into the video, and it leaves AI detectable visual cues for 10 minutes, the AI will never find it using binary search. It will skip to 30 minutes, see nothing, skip to 45 minutes, see nothing, skip to 52:30, see nothing, skip to 56:15, see nothing, and fail at some point when it can't divide the video further. Binary search would fail in this scenario. It's not just useless, it's an abject failure, and the AI was a waste of processing power when you could have scrubbed forward five minutes at a time instead. That would have found the visual cue, but would not be a binary search.

                    • The time of the event doesn’t necessarily coincide with any of the times that you’re checking. That’s the whole point of looking for visual cues.

                      But one of them potentially will though. A gun firing leave smoke behind.

                      Even if there's other smoke in the video, you're looking at 5 minutes of a 24-hour video, and not scanning through 24 hours of a video manually. And an AI could use a binary search to find any moments of smoke (or not). Not saying it's a one-size-fits-all solution, just one very important tool in a toolbox.

                      I don't mean to be rude, but I'm exhausted talking about this topic, and so if you don't mind, I'm just going to bail at this point.

                      Thanks for keeping it civil.

                      • If we're talking about a 24 hour video, then we definitely can't find every instance of smoke. If there isn't any smoke exactly 12 hours into the video, then it throws away the entire first 12 hours. Any evidence that could have been found in those 12 hours is gone. A binary search throws away half of the information at a time. It super can't locate multiple instances of something happening.

                        I've been wrong in arguments before, it feels awful. The best things to do are either address the misunderstanding in the original comment, or not engage with anyone else who feels like arguing more. One thing I miss from Reddit was being able to toggle notifications on a per-comment basis.

365 comments