Name: Motion Detection Algorithm Edits my Videos
Uploaded: 2020-03-27T18:24:14.000Z
Duration: 10 min 49 s
Description: Thousands of YouTube videos with English-Chinese subtitles! Now you can learn to understand native speakers, expand your vocabulary, and improve your pronunciation...

Past simple

And this is like super rough brainstorming.

The imperative

Don't take this as fact or anything, but let's compute the variety, so compute the difference from the current frame to the previous frame.

Like Imagine if we had an array that would tell us how different this frame is from this frame.

If you haven't seen the first video I made on this project Go on.

Watch that, because this is gonna build on top.

But, you know, when I was doing the random edit thing, it looked pretty good, but it was still kind of random.

It would be much better if my algorithm was able to actually pick the frames that have the most the most difference from everything else.

Now, how are we supposed to calculate that?

Well, all I was thinking is we We look at every single pixel and we just compute the difference.

And if we do that for every single frame in the video, maybe I'm not saying it's gonna work.

I know we're gonna be normal and hang out with some friends, and I think they want to go to a club, which is not really my thing.

It's a few days later the Sunday we tried to run the first test of our variety algorithm.

Picks A format must be this so it looks like this can convert like, one pixel format into another.

I just I don't know how it works, So I'm gonna go look it up Way realized last night that we want to do our calculation and RGB We tried to convert the RGB and now we get these super weird errors and segmentation faults and it doesn't look like it completely broke.

Oh, dude, it's almost like almost real time.

It's It's slower than real time, a little bit so we can see it looks like I've moved the camera here, and then here is like, I haven't really moved the camera.

Let's see, when we change clips 0.5 point 2.3 The next clip is 0.23 because it's a new scene, so we can actually mathematically detect like we think that there's a new scene here because this frame is so different from this frame.

When I hit, play here at 40 seconds is right.

So this is this is this right here and then it goes back down.

And then, Ah, at about what is that number at 59 seconds.

So this right here is detecting the movement for me, right?

When you code for two hours and then you run it for the first time.

I've spent so much time working, and I feel like I haven't explained what's going on, so I'm just gonna do a full run down.

I tried to write this function that would be able to tell the difference between one frame and another frame.

All I have to do is go through every single pixel and subtract the red.

If you're If you're talking about RGB, you know, subtract the Red Channel, subtract the green Channel, subtract the Blue Channel like you could see the actual code here.

That how simple it is because we're dividing it by 2 55 This will give us a number between 1.0, and 0.0.

If the two pixels are completely opposite, like black and white, then we would get a one point.

Oh, so we basically get the average of every pick Soul between 0.0 on one point.

Oh, for the Red Channel and the Green Channel and the Blue Channel.

And then we get the average of all three channels so you can see there's no crazy math or anything.

We compare this frame to this frame and then we shift.

We compare this firm to this frame and then we shift.

And what we end up with is an array of values between 0.0 and 1.0, on that represent how different this frame is from the previous run.

The very first demo off this this program, this is the output file.

So that means any slight bit of emotion it's gonna pick up and make that cut around this one could be any length because it depends on how much there is in the video.

So, yes, we have a loading bar on command line, so we'll say in code new file.

Dude, you just cut straight multi view, bro.

Ah, I'm doing the tests on the same data.

If you can remember, we're going to say on Lee detect things above 10% motion.

So we're just gonna be sitting here for a while.

Basically, I just two hours an hour and 1/2 of sitting here, and then we get a Sega fall.

We're gonna try and do sequence to again.

The one that gave me a SEG fault last time.

I'm gonna leave this running overnight, and we're gonna come back on, see what's up.

2.7 hours of video and the output is like 100 seconds of video 100 seconds of video so you can see it says if we cut 98.9% out.

However, this is not gonna be like a good edit.

This is just gonna be a bunch of different scenes together.

So I want to see if the algorithm is capable of identifying what it thinks is different scenes.

The audio's I'm gonna meet the audio, but let's is going to be first time basic scene detection, See what it did.

Honestly, dude, I'm kind of surprised how good it looks.

It is detecting a lot of motion, but it is cutting the scenes that look different from each other.

Now this algorithm cannot tell if frame looks good or bad, But what it can do is compare the difference between all these frames and trying to pick the most different ones.

I want to show you what the original and it looked like at this part.

So the one on the right is the algorithm.