Creative Planning Meeting – Vexnomics Offices

The team are excited to try TikTok, but they’re not sure what videos to create. We need to know what works in our industry, so we can at least get to good enough. There is the TikTok ads library but it’s a lot of manual work to go through those videos…

Why don’t we use Google’s Video Intelligence API? 
It’s actually really good, and won’t take more than a few seconds per video, for short videos like this.
It’ll spit out the labels of what it detects is in the video
Then we can pivot them by label to see which labels are most popular
Shall we try it with the top 20 videos?

Computer vision has progressed exponentially in recent year, to the point where we now have the processing power to dissect videos frame by frame and detect what’s in each shot. Machine learning algorithms are trained on millions of videos and their labels, so that they can learn to recognize objects, faces, and landmarks. The technology can be used to automatically tag videos with metadata, so that they can be easily searched and indexed. There are a range of applications from security and surveillance to image and video search, and decomposition analysis: breaking the best ads in your industry down into their component parts to analyse what’s working.
Companies like Google have democratised these techniques through open source libraries and commercial APIs bringing the cost of experimenting with them down considerably. Now anyone can tag any image they like in a few seconds without owning a supercomputer or knowing anything about machine learning. Marketers who want to go a level deeper with industry analysis, can use video intelligence to automatically tag each video in their category, and then pivot to see what labels are present across most of the top performing videos. Then these ideas can be used as new hypotheses to test in your creative testing plan, to validate if they work. This approach will help you at least get to ‘good enough’ and serve as a foundation for further testing.

Video Analysis

Hey, let me show you how to do a video analysis on your industry. So one of the really cool tools is available to everyone now is from TikTok that adds library and Facebook or Meta.
They have an adss library as well. But I think the TikTok one is really great and it's especially interesting that it's video.
So you could go through and, you know, you can filter for your country you can filter for your industry. In this case I'm in the marketing and advertising industry, so that's what I care about.
And then you can even filter by objective, in this case, conversion rate, and then choose, you know manure number of likes or, you know, duration, et cetera.
You could also search for specific products. So really, really comprehensive. When you click into one of the videos, it actually shows you some intelligence about, you know, where attention was in the video.
You know, when do people click, when the conversions, et cetera, which is pretty cool. It also tells you how many likes or comments or shares it got.
And then you can watch the video as well. What I seeing the benefit. So you know, it's, you could really just do this manually.
You could go through and watch every one of the videos in your category here. It's like sorted, descending by reach.
And and you could do some analysis. So you could make some notes and say, Okay, these are the sorts of patterns that I'm you know, that I'm, I'm noticing, but obviously there's quite a few of them.
And, you know, you don't necessarily want to do this you know, manually, right? That's watching a lot of videos. Maybe you could get an intern to do it, but, you know, are they really gonna pick up on all the different labels that you want to tag these creators with?
And the, the reason you're doing this is pretty strategic, right? Like, if you are gonna start advertising on TikTok, you wanna know what's working in the category, and these are the top, you know, top percentage of people in the industry.
So you really wanna kind of do this analysis in a more automated way. Okay? So so this is how it works.
You can go into Google CoLab and then you were using this Google Cloud intelligence here. This is obviously like a video intelligence tool.
It's similar to the image tagging that they do. If you hit play here and run this then it will just download everything into this Google CoLab.
And, you know, don't need to have it on your local computer, and then you just click run reset runtime here.
Cool. So there's a few things you need to get set up. So one is you need all the videos. I've given you a folder here.
I just went through and downloaded them all the way that you download them. And, you know, this is a case of like learning from the you know, don't, don't necessarily worry about the method, right?
Like learn from the approach rather than the specific method because the way that I downloaded these videos might not work in the future if TikTok changes things.
But rest assured that basically if there's any video that's playing on your screen, you can, you can get that video and download it somehow.
It might be against the terms of service. So you know, I wouldn't necessarily recommend doing this as, you know, use your own best judgment.
I think in this case you know, I'm not creating a product from this. It's literally just, you know, teaching you how to do video analysis.
You could get your videos from anywhere, you know, could be just you know, noting them down. You know, when you when you see these ads in the world whatever, whatever it is you want.
Cool. So the way that I got the videos, just so you can see, and this works on lots of different sites.
If you get to network the network tab, so I, you know, right click and then you go to Inspect Element, and then you get to network tab, and then you can see basically any requests that come in.
And I'm just searching for mp4 here. And what ha, what happens when I click the video? When you start seeing the benefit?
Okay, did you see that come in? So this is a request that to sent to its, you know, main headquarters and said, We need this video.
So you can just copy that URL past the, and there we go. Now we have it in a different tab.
And this is from Akamai. This is a website that basically hosts the video for, for YouTube. So if I right click, I can now say video as, and then you know, and then I've got it here.
Cool. So that's how I got those videos just in case you want to do the same for your own category.
And you could note down, for example, the number of like cigar or comments or whatever. You know, it's really up to you how you want to kind of associate these.
But in my, like, in my opinion you know, these are all like, should be like pretty good if they've, you know, the top of the industry.
So I'm just, I just grab the, the top 20 videos and then I uploaded them all here. So you can see them all over here.
You can download those two if you wanna run it. The other thing you need so just gonna, you know, run that.
And then if, if you've done this correctly, then you should see the list of videos update here. It's just basically finding anything that ends with mp4 putting it out.
So you know, you can actually display a video here. So this is just the code to display. I'm just displaying the seventh or eighth video in the list here.
And if you run that, you can see it's, this guy just takes a little bit of time to display, cuz you have to kind of compress it and base 64 encode it.
So just wait for that might be quicker just to click <laugh> to see it and you'll look, repeat it. But while that's running the, the other thing that, oh, here we go, it's worked.
Cool. So now you can play Reasons why your stores stuck below a hundred K a month and you're not seeing results like this.
Cool. So the the next thing we need to do is go through the follow up the, the setup instructions. If you go to this url so this is for the the api.
You'll need to click enable api and then you need to grab a service account. So if you follow that link it should help sort you out.
And once you've done this, you got the service account you just upload and then rename it to service account J and then that should work.
Cool. So this is how the video intelligence works. You don't need to worry too much about what this code does.
I'm just taking that same video with that guy. And then displaying the name, opening them, opening the video locally and then sending that operation out to, to Google.
And then Google's gonna send the labels back and I can loop through them in order to display them. So let's try this, a processing video for label annotations.
So it's up to, yeah, this part here, and then it'll say finished processing, and then we should start getting the labels.
It takes a few seconds in in my experience like, depends on the video. This is a pretty short video so, you know, it won't take very long but they'll start to display them.
While, while we're waiting for that, I'm just gonna run you through what this does. So this is basically doing the same thing, but doing it for every video that we've got.
So we've got 20 here. I mean, I dunno what the limit is. Maybe you could do hundreds in CoLab. If you're gonna do a lot of video texts, I recommend actually getting a way to like upload them into Google Cloud storage first, because you can do this automatically in Google Cloud storage rather than downloading it on your local machine.
Cool. So this is all the labels and this is what you get back. So there's different types of labels you can get, but I found that the shot labels is the most interesting.
So what it does is it kind of breaks the videos into different segments. So when, when the screen changes from one to another then that's a segment.
So in this case, it's like between 7.8 and 10.87 seconds as a segment. And it's got 30% confidence that there's a font on the screen at that time and 50% confidence that it was on screen between 56 and 56.5.
So you could be clever here and you could get rid of the ones that have low confidence. I'm just you know, taking all of them for now.
I just wanna kind of get a general idea of what works. So you can see head animation, human text, graphics, eyebrow, forehead, hair, facial hair, chin, cheek.
So tell me that like faces you know, showing up a lot in this video. So that's interesting. Cool. And you can run this and this, this took, took I think about like six to 10 minutes.
And then once you have all the labels you can run this and then this is just gonna pull them all together into a spreadsheet, right?
So this is a data frame, pandas data frame. And now we have, you know, for each video we have one entry for each label.
So you can see like sometimes the videos don't do much like this one, only recognized hair and long hair, which isn't great.
But then some of them have lots of different interesting labels. So this one, they f they figured out they had a L E D back lit LCD display, which is pretty interesting.
There's also a computer terminal display device numbers they noticed, numbers cetera. Cool. So now if we, you know, it's not that easy to analyze here.
So what you wanna do is kind of group them by the the labels. So what this does is it basically joins it together.
So now we have it by label and then how many videos it's in, and like, which videos is it in?
So this is you saw for a paper trail, like if you wanna go back and see, okay, what specific video did it notice album cover in you can do that.
And then once everything is in that format, then we can do this which is a filter, and it just filters it by anything where it's in at least three videos.
And, but it's not in all of the videos because, you know, if it's in all the videos, then it's too common to like care about.
If it's only in three of the videos then you know, if it's like in less than three of the videos then it's, you know, probably gonna be you know, not, not that useful to, to know, right?
It's, it's like too rare in in this top 20 list. So I'm just filtering that way, and then I'm, as I'm sorting it by descending.
So ascending equals false. And you can see that 13 of the 20 videos have texts and them and nine have font, seven have logos, six have brands, and there's graphics graphic design.
So that immediately, like, it's this type of analysis, you know, it's different every time, but this tells me that in marketing and advertising at least most of the ads have like, text overlaid really important.
So that's something that you know, I would only get right when I'm making a TikTok ad. And then it looks like you know, a lot of these have like logos, brand graphics so maybe investing in you know, overlaying some custom graphics in the video could help it do well.
There's also a smile, so we noticed a lot of facial features and some of them so it's worth kind of looking at that.
And again, these are just cursory elements that like, interesting for analysis. You know, you, you, this isn't a guarantee that if you put text in your video, it's gonna work.
But it's just showing you what are the common features, Like what are the memes like what are the individual components of the video that tend to show up for the best performing videos.
You could also do this potentially for the worst and, you know scroll all the way down and see, you know, which, which videos and not getting much traction.
And that might tell you what to avoid. So you know, it's really up to you how you use this type analysis.
I'm just kind of teaching you what's possible and then you can adapt it for your own purposes and run this at the end and it'll download the video analysis file which shows you, you know, basically this spreadsheet.
You can use that in Excel and do some further analysis afterwards. Okay. Hopefully this was useful for you and you know, it's it's a pretty cool tool that you can use and this AI is only getting better, so it's going to, you know, detect more and more interesting stuff.
All right, Thanks.

Why might you want to work with AI to tag your videos?

Which creatives feature the label “Animation”?

What’s the most common tag in the featured videos?

Why does it make sense to remove tags that don’t appear in at least 3 videos?

Why might you have to be careful about making decisions solely off this type of creative analysis?

Before you can build great ads, you need to get past average. The first thing I do is swipe a collection of the best videos in the industry (i.e. from TikTok ad library), and label them to identify what they have in common.

Vexpower Learning Certificate of Completion

Congratulations fabulous-drone-53 💪