Discover the cutting-edge world of energy-efficient machine learning, edge AI, hardware accelerators, software algorithms, and real-world use cases with this podcast feed from all things in the world's largest EDGE AI community.
These are shows like EDGE AI Talks, EDGE AI Blueprints as well as EDGE AI FOUNDATION event talks on a range of research, product and business topics.
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
0:00
|
15:01
You shouldn’t need a warehouse full of broken products to build a reliable vision model. We dig into how synthetic data flips the script for manufacturing and quality control, showing why controlled scenes, perfect labels, and rapid iteration can outrun old pipelines of manual collection and error-prone annotation. With Sherry List and Goran from Synthetic AI Data, we walk through the strategy behind a no-code engine that empowers fusion teams—engineers, developers, and operators—to generate the exact scenarios their models need.
The story centers on a high-stakes, high-volume domain: metal and aluminum cans. From standard to sleek, slim to stubby, tabs in different materials and colors, and both beverage and food lids, we map the defect landscape that actually matters on the line. Bent, broken, missing, and lifted tabs are recreated with photoreal materials, realistic reflections, and varied lighting, then captured from top-down and side angles to match real inspection setups. The result is control—over class balance, severity, camera pose, and environment—so teams can stress-test models and discover what improves accuracy before hitting production.
We also reveal the scale required to make a difference: roughly three million synthetic images, high and low resolution, fully annotated in COCO, YOLO, TensorFlow, and more. Releasing the dataset under CC0 via the EDGE AI Labs removes friction for researchers and practitioners to explore defect detection, domain randomization, and multi-view training. Along the way, we share a cautionary tale of “screwdriver datasets” and explain why simulation delivers safer, faster, and more reproducible results. If you care about computer vision performance, cost, and time-to-value, this conversation offers a practical blueprint you can use today.
Subscribe for more deep dives, share this with a teammate who wrangles labels, and leave a review telling us what you’ll build with the dataset.
Hi everyone, my name is Sherry List, and I'm co-founder and CEO of a startup called Synthetic AI Data. So judging by the name, we are actually a synthetic data provider for training computer vision model. So we know the pain, and then we decided to come up with some cure for that. We are based in Copenhagen, Denmark, and we actually that we are we have a platform, we have an engine that we are working on it for about uh three years. And um our startup is um kind of backed by Microsoft Ware startups and VIDIA inception program, and also we got a grant from Innovation Fund Denmark, which is uh like a um uh governmental grant from uh from a country.
So uh we also um last year we won a Microsoft uh partner award in a category of um startups uh because of our innovative way to solve uh this problem. So um about our actually um our solution, uh as I mentioned that we actually that we have um we have a platform, we have an engine that uh with our engine, our customers, they are they can they can create the scene that they want in our platform, and then um they can easily without uh any hassle train uh actually generate the annotated data. So uh we chose the approach to the approach to make sure that our platform is a no-code solution, so anyone can actually use that uh to generate the data. Because we believe that um in fe in future, actually not even in future, but currently all of our teams they are fusion teams. So it consists of people with a with a technical background, with a developer background, and also business background. So we should make sure that everyone is able to uh is able and empowered to actually that to uh to generate this data. And of course, we all agree that uh AI needs a lot of data. And currently, the way that people and all of us we are generating the um data for training AI is that we need to spend so much of a time to collecting the data. Once we collect the data, we need to annotate it, and then that annotation and collection is time consuming, and it's a very most of the time it's a very manual process. And we all know that as human we make mistakes, and in this case, this mistake costs us uh the accuracy of uh our model. And uh that's why that we believe that the solution for that
is synthetic data. So synthetic data is the data that is generated by simulation, by computer algorithm. They are very much like real-world data, uh, but they are uh in reality is they've been synthesized. And the good thing about it is that since it is generated by computer, it's all it's out of the box comes with all the annotation data. So you can have the segmentation, you can have the all the uh points for the uh for the object detections and so on without doing any extra work. And um also, as you can see, that for example in this picture, uh since this scene is generated by um actually that is a is a synthetic scene, so you can have even more than one object in each picture. So you can fine-tune it because it's your uh because it is this is what you actually that you have all the control over it, so you can have all of this annotation out of the box. And um one thing is that we are not the one that we say that, hey, synthetic data is the future. So Gartner is also saying that by the 2030, almost all the data that is being fueled for for AI model is being uh is actually all of this data that we use for training the model is fueled by synthetic data. And uh so in reality, you cannot have accurate data and develop perform data without using synthetic data.
And in our actually, in our company, we have um our focus is on uh mostly defect detection, because we know that that's an area that uh most of the uh actually that the the industry is uh lacking the data. So just imagine um that, for example, if you want to make sure that you want to train a model for uh finding the defect in this clicker, how much of a different types of defect this clicker could have. And then how much of a time you need to spend in order to create this defect to and take a picture and then annotate the picture to be able to have a welfare-performed model. And that is why that we chose to actually that to have this as the focus of our um our company. And I'm going to hand it over to Goran to actually show us uh some of the recent uh project that we did.
Hello, my name is Goran, and um I'm co-founder and city of uh Synthetic AI Data. And uh recently we became uh strategic partner of HAI Foundation. That's why we are here today. And um together with their uh data working group, we discussed some use cases like what problem in industry we could tackle, and uh we decided to go after this um problem of metal and aluminium cans. We all know how they look like they are for beverages and food, and um they are used nowadays uh in industry a lot because they are easily transportable, uh they are durable and uh highly recyclable, like uh 90 to 95 percent of these cans is recycled. The thing with with the cans uh that you maybe don't know, it's a huge industry. Uh there is 627
billion units produced each year, and uh that approximates into 1.72 billion cans per day being produced out there. So it's a pretty huge industry. And um just in uh US uh in 2020, it was produced 120 billion of cans. Um in Europe there is something around 100, China is 75, and so on. So uh huge global industry, and projections are also that this will grow from 120 to 173 by 2030. This industry is um yeah, steadily growing over the years.
Um challenges with metal and aluminum cans in the in the production is um that those cans could have the broken tabs, could have missing tabs and lifted tabs, and this is basic functionality that we actually need to be able to open the can, right? So we want to be able to detect is this can correct or it's not. For this purpose we decided to develop a dataset, and I will show you through slides some of the pictures from from this dataset. Um here what you can see is um standard beverage can. Imagine cola or beer, that those are the cans. But if you look a bit closer, like those cans are a bit different, depends on the on the top cover, right? And they also have the different opener tab on top for all of them. They come, this is as I said, standard one for the drinks. Then there is a slick version, which is used for energy drinks mostly. You'll probably also see those a lot. Then there are slim versions, then there are stubby versions. All those scans come in many different materials, and tabs come in many different colors. So here showing you some of the variations. And what you are looking over here are synthetic images generated by simulation on our platform. This was for standard, then uh also some preview of variations for the for the slick version of cans, slim, stubby, a lot of a lot of those variations. Uh in this data set, as we were saying, we are focusing a lot on the defect detection and simulating actually the defects. And here you can see one example of a defect, like this is correct picture, and on the right, there is a defect, right? This tab is bent. This could happen in in uh production, and this is something that we want to detect.
Um it might sound easy if you haven't worked with the computer vision, like how to test the defects and such, but I can give you one example of one enterprise company that had a model in production not for this purpose, but for their product being uh able to detect the defects. And the way they trained their their model, they took they took a lot of um a lot of their products, their data scientists took screwdrivers, damaged those products, took pictures, and then trained the model, right? We don't want data scientists to run around with screwdrivers, and we don't want them to break down our products, that's not sustainable, so it's much better to run simulations and create the defects. Uh few more examples of defects. Also, example of a bent tab, another variation of a bent tab. A few more previews here from the from the different versions. Um this is uh preview of a broken tabs that can also happen. So in the dataset, there is simulation of the broken and bent tabs, as you can probably see. I hope that you can like tab is half broken over here. In the dataset, there are beverages, there are food cans also. Food cans come in many different shapes, they also come with different uh tabs on top, they come with uh in uh different materials, tabs can be in different materials and so on. During the generation of data set, I was showing you now just uh top view and one orientation of the can. During the generation, we generated uh images from many different angles, rotating those cans. And uh maybe pictures are too small over here to be visible, but you'll have access to that data set. Feel free to check it out. There are different lightnings and shadings on those images, so light reflects uh differently on those pictures from the can since we have the metal surface, and also the the shadows are falling differently. In um production, we could inspect the cans in two ways, with camera from the top looking down, 90 degrees, right? But we could also inspect it on the side because from the side it's easier to catch some defects such as can being bent. So uh all those synthetic images are generated with uh also in a different view with uh camera angle, and here you can see that those tubs are tubs are bent. Maybe this one would be easy to detect from top, but this one probably wouldn't because it's not bent so much.
Where we ended up is uh generating the world's largest CAN synthetic data uh synthetic images dataset, uh resulting in 3 million images, 2.985600 images, uh 267 gigabytes. Uh this dataset will soon be available at the HAI Foundation uh AI labs, so you can go over there and download the dataset and inspect it. We look forward to see what people will build on top of it. Um all images are available in high and low resolution. Uh they are fully annotated in uh Coco, IOLO, TensorFlow, and uh some other formats, so being uh ready to be used for training the model, and they
will be available under the Creative Commons Zero license so you can build applications on top of it. Um yeah, so um I would wrap up here and would be happy to hear your questions.
