Eric Cheng, co-founder and CEO of Intrinsics Imaging discusses his journey from developing critical vision systems for the US Navy to building advanced AI inspection systems in the cloud. We discuss the benefits of Intrinsics’ approach and the comparison to traditional computer vision.
Quickly discover & access sections of the transcript by clicking the arrow:
- Introduction & Guest Overview Host Josh Eastburn introduces the episode and welcomes Eric Cheng, CEO of Intrinsics Imaging, to discuss how cloud-based computer vision systems are transforming manufacturing.
- What Intrinsics Imaging Does Eric explains Intrinsics Imaging’s mission: delivering scalable, cloud-native computer vision systems that blend traditional CV with modern AI for real-world industrial use.
- The Case for Cloud-Native Vision Systems Eric describes how moving vision workloads to the cloud enables faster iteration, centralized management, and better use of AI without compromising factory performance.
- AI + Classical CV: A Hybrid Approach Rather than choosing one or the other, Intrinsics Imaging combines traditional computer vision and deep learning — leveraging each where it’s strongest.
- Learning from Real Deployments Eric shares lessons from real-world customers and how Intrinsics uses feedback loops, retraining, and usage data to continually improve system performance.
- Business Impact & ROI for Manufacturers Eric highlights where the biggest cost savings and efficiency gains come from — including faster deployment, reduced errors, and simplified system updates.
- Future of Cloud-Based Vision in Manufacturing Eric offers his outlook on what’s next: from edge-cloud collaboration to plug-and-play models that radically lower the barrier to adoption.
- Final Thoughts & Where to Learn More
Episode Transcript
[00:00:00.520] – Eric Cheng
“Having worked in periscopes and body cameras, we were used to dealing with not the best hardware and very difficult environments. We were used to having to make the software really good.”
- Introduction
[00:00:15.560] – Josh Eastburn, Host
Hello and welcome to the MV Pro podcast. Today, we’re picking up a topic we introduced a couple of months ago in episode 007, evaluating AI vision technologies through the lens of system integrators. The cutting edge of any field often falls victim to hype, so our hope is to present a grounded perspective on what is really happening.
What’s possible now? How well does it work? What pitfalls are there to watch out for?
2. What Intrinsics Imaging Does
[00:00:37.760] – Josh
Today’s guest is Eric Chang, co founder and CEO of Intrinsics Imaging. Intrinsics develops specialized computer vision algorithms to automate repetitive vision tasks in manufacturing. It operates and maintains hundreds of cloud-based vision analytics deployed to manufacturing facilities worldwide, leading to reduced downtime, waste, and claims. Eric leads the development of computer vision and AI systems for Intrinsics and has over two decades of experience developing mission critical vision systems for some of the most challenging environments on Earth. He began his career developing and deploying numerous algorithms to enhance situational awareness and automate repetitive tasks for periscope operators on US Navy submarines. After a decade in defense, he co-founded Intrinsics Imaging, which developed vision solutions for numerous law enforcement agencies around the world, including the LAPD. Following the acquisition of its law enforcement business, the company refocused on manufacturing, where for the past seven years, Intrinsics has helped modernize industrial inspection through intelligent imaging. Eric holds a bachelor’s degree in physics from Harvard University and a master’s in electrical engineering from USC. Please enjoy our conversation today.
[00:01:45.120] – Josh
Let’s jump into the beginning. Eric, you spent 10 years working on vision systems for Navy Marines. What was that experience like for you?
[00:01:53.920] – Eric Cheng
Well, Josh, it was a great experience, actually. I’m quite grateful that I was able to take that technical field of study that I love, which is computer vision, and apply it to such an interesting, important problem so early on in my career. Periscope operators have an extremely difficult, mentally taxing job. You can imagine they have to maintain the situational awareness of the submarine. They’re just a couple of feet below the surface and just looking through a little tube, basically. They have to spin that tube around to try to understand everything that’s going in that full 360-degrees field of view around them. They have to do that in extremely difficult environments. It could be nighttime, it could be really bad weather conditions, rough seas, it could be hazy, foggy, dark, and you have to keep track of everything that’s going on around you. It could be a lot of different surface ships out there. It could be things flying in the air. Our job was to basically make that video as clear as possible so that the operators could see what was going on. Then also we started to automate some of the repetitive tasks that they had to do, keeping track of the different contacts that they saw out there.
[00:03:10.700] – Josh
Interesting. I imagine that a lot of the algorithms that you developed for the Navy are pretty hush-hush. You’re probably not at liberty to talk about a lot of that.
[00:03:19.160] – Eric Cheng
Yeah, I can talk in generalities. It’s fine.
[00:03:23.150] – Josh
I’m really just curious about the part where you said things flying through the air. I’m like, is that a helicopter or is that missiles or both?
[00:03:31.120] – Eric Cheng
Whatever it might be, Josh.
[00:03:32.440] – Josh
Whatever it might be. Fantastic. So juicy. That’s enough of a teaser to get me pretty excited. No, that’s fantastic. That’s really interesting. Then, of course, what I’m interested in is tracing your own trajectory. How you got from that to what you’re doing today, which is focused more on manufacturing machine vision, correct?
[00:03:50.920] – Eric Cheng
Yeah. I actually met my Co-Founder, Nick [Nicholas Flacco, ed.] doing the periscope computer vision work. He was a former submarine captain himself, and he knew all of the challenges that the operators were going through that we were trying to solve. I worked with Nick for about 10 years doing that. Then at that point, we got interested in turning our focus a little closer to home, where we had a more direct contact with the end user. That got us looking towards a natural extension of the work from the military over to law enforcement. We were based in Los Angeles, and right around that time, the LAPD was rolling out something like seven or 8,000 body cameras for the entire police force. We saw an opportunity there to take the computer vision expertise we had developed for periscopes and apply it to all that video that was being generated with body cameras. That was the initial focus of Intrinsics when we spun out and started that company. We worked with law enforcement for, I want to say, three or four years or so before that technology was acquired. Then after that acquisition, we transitioned over to manufacturing.
3. The Case for Cloud-Native Vision Systems
[00:05:16.440] – Josh
That already gives you a pretty different pedigree from a lot of the other vendors in this space. You chose a different way to put together your offering, right? So Intrinsics offers cloud-based vision systems. What led you to make that decision?
[00:05:34.100] – Eric Cheng
Yeah, so that was somewhat a combination of happenstance, serendipity, and our natural expertise as more of a software-focused group of engineers, like you were saying. We were really coming at this industry more from the computer vision and algorithm development and software side of things rather than the hardware side of things. Having worked in periscopes and body cameras, we were used to dealing with not the best hardware and very difficult environments. We were used to having to make the software really good to handle all of the complexities of dealing with those difficult platforms. We really came at this from more of a software perspective. Then when we were ramping up and starting to gain traction in the manufacturing industry, that was right around the time of COVID. Just because of that, we really didn’t have the chance to actually go in person and visit a lot of these manufacturing facilities. Just out of necessity, we had to just ship them the camera, the hardware, have the plants, resources, mount the cameras, connect them to the cloud. And then once the cameras were streaming to the cloud, we could do what we did best, which is just take a video feed from wherever it was coming from and to develop the software to monitor that 24/7.
[00:07:04.920] – Eric Cheng
That’s really what led us to that cloud model initially. Then we found that we never actually needed to go back to the plant after that. Once that video feed was connected to the cloud, and we were able to monitor it remotely. Actually, it offered a lot of advantages. It allowed us to keep an eye on the performance of the analytics and the algorithms in real-time, 24/7, and make fixes remotely and continuously and instantaneously.
[00:07:36.480] – Josh
That was a combination of your expertise, but then also the field experience of trying to bring these systems together that led to that insight of like, oh, maybe, wait a second. Maybe we don’t need to take the traditional approach to this.
[00:07:51.000] – Eric Cheng
Yeah. And what we found initially was that a lot of times you may have a nice sophisticated algorithm that works off the shelves. But then once you try to go and apply it in a particular location or a particular site, there’s always a lot of massaging or optimization or tweaking that has to happen to get that algorithm to truly perform well in that particular site. We work very iteratively. We get something out there processing the video feed, and we’re constantly looking at how it’s performing, maybe what false alarms it might be generating, and then making tweaks and optimizing the algorithm until it achieves the desired level of performance. That continuous improvement process really works well with a cloud-based system where we can just continually update the software remotely as needed.
[00:08:52.740] – Josh
I like your point about the degree of specialization that you’re able to get into. I think we’ll come back to that. But I’m also wondering, do you find that that leaves a lot for users to do who are in that last mile of the network or people who are on site. How does that not become an obstacle for you? How have you dealt with that?
[00:09:13.420] – Eric Cheng
Yeah, somewhat unexpected, it’s actually the opposite. Once the plant resources have installed the camera and the video is streaming to us in the cloud, basically, we take over from there. That allows us to just remotely monitor the poor performance of the analytic and then make adjustments as needed. What we find is that a lot of that last mile work to get the system truly working and working under all conditions, a lot of that is really software-based work where you’re trying to make tweaks and improvements to the algorithms to handle unforeseen complications in that environment. A lot of that work previously somebody on site who’s maybe not necessarily a vision expert would have to do to train the system themselves. I think that our customers are finding that they would prefer to let us do that if we have the access. Interestingly enough, in the six or seven years we’ve done this with hundreds of cameras, I don’t think we’ve had a single camera failure. The cameras are quite robust at this point. If something does go wrong, we can ship out a replacement real quick and then make the necessary tweaks on the software again to adapt to any thing that has changed and then get the system back up and running.
4. AI + Classical CV: A Hybrid Approach
[00:10:41.700] – Josh
Interesting. Okay, this is already pretty exciting because you’re providing a different lens on how we typically approach problem solving around quality inspection. But AI is also something that you have stated is very core to your offering. That term gets batted about a lot today. Everything has AI in it. Obviously, as technology people, we understand that that means different things depending on what product, what tool we’re talking about. In the realm of machine vision, When you talk about using AI or AI vision algorithms, how do you distinguish that? What technologies are you talking about that distinguish that from traditional types of vision technologies?
[00:11:26.400] – Eric Cheng
I would say that the distinction there is that traditional computer vision approaches are generally things that you directly program where if you have a pretty well-defined defect that you’re trying to detect, let’s say it’s a hole or a stripe of some kind in a particular location, usually you can directly program an algorithm to look for that specific thing, and you can give it some parameters and different thresholds and things to optimize the performance of that. But you’re directly coding in the logic of what you’re looking for, how you’re looking for it, and what records to use to a LAR. Now, that generally works for very well-defined defects like, Hey, there’s a hole I’m looking for in this spot, or I’m looking for this line. It needs to be here, and it can’t be more than three inches off the edge of the sheet, or something like that. But where that becomes more and more difficult is when you’re trying to do slightly higher levels of abstraction. Like, show me this picture and tell me if you would consider this to be a piece of debris or a lump or whatever it might be, right? Those are slightly more abstract concepts, and I would call those more like image recognition or image classification type tasks.
[00:12:49.300] – Eric Cheng
And that’s really where the AI approach excels. And so that approach is more of basically a machine learning approach, where instead of directly programming in the algorithm the different rules or the different logic that you’re going to use to detect something, instead, you’re taking training data, examples of the different types of things, let’s say they’re dense or scratchy or something, and usually taking label planning data, feeding it into some machine learning algorithm to train an AI model or a neural network. Then that network then becomes very good at mimicking the human’s ability to quickly label an image as this or that.
[00:13:34.920] – Josh
Okay. The intense pattern recognition. What is the net result of what that allows you to do that you wouldn’t be able to do with that traditional approach?
[00:13:47.080] – Eric Cheng
Yeah so, it actually allows us to do a lot more things that more closely mimic what humans would do. We like to say that anything that a human can see in the video, we can usually train a computer vision analytic to see the same thing. In some cases, we will use traditional computer vision approaches if that’s a fairly simple thing you’re looking for. But as you get to these more subtle things, that’s where we would start to apply the AI-based approach in order to mimic human performance. This basically allows the customer to dream up more and more applications for machine vision and how that can improve their production process, because they can just start to imagine, Where along the production process do I wish I could have somebody standing 24/7, looking at it, not getting tired, and telling me when this particular condition happens and then taking some automatic response to that. That’s obviously very difficult to do and mentally taxing to do for humans. But if a human can do it, you can usually frame the AI model to do the same thing. And so that could be anything from looking at a production line and trying to detect when maybe there’s some jam up or build up of products that could potentially cause some downstream problems where they have to shut the line down and take a couple hours to clean out a piece of machinery, or it could be looking at packaging of the finished product on its way out the door and trying to mimic what the manual inspection would be doing in terms of trying to verify the integrity of the packaging.
[00:15:38.580] – Josh
Where do you think the right line is between when you apply these contemporary AI techniques and where you apply the traditional imaging techniques?
[00:15:49.580] – Eric Cheng
Yeah, that’s an interesting question. I think it’s generally better to use the AI techniques when it’s a fairly unambiguous recognition task that most humans would agree on the answer to relatively quickly with not too much context. If you took a picture of your defect and you showed it to three operators, and They all within a couple of seconds could tell you, Yeah, that’s a scratch versus that’s a dent or whatever it might be. Those are great tasks for trading a model for because the agreement between the operators and the lack of ambiguity in terms of labeling that imagery is really what allows you to generate a lot of training data. Where the traditional techniques are often better is when you’re looking for lower-level things that are much more well-defined and easily defined. So if you’re looking to, like I said before, identify the presence of a line or a shape, and tell me basically where it is, if it’s there, and then if it’s there, tell me how strong it is, like how brighter. What the contrast is. All those things are generally better suited for the traditional computer vision approach, where you can actually start to pull out some of these measurement values as well, rather than just jumping straight to the answer of, is it done for a scrap?
[00:17:23.660] – Eric Cheng
Okay.
5. Learning from Real Deployments
[00:17:24.520] – Josh
I guess you’re touching on some of the limitations, too, of AI, which is that if you get three operators together and you show them an image or a series of images and they can’t agree on what it shows, you shouldn’t expect that the AI is going to be able to tell you, right? Because it depends on you being able to train it. Yeah?
[00:17:43.020] – Eric Cheng
Yeah, for sure. I think that’s one of the unexpected challenges we’ve come across, or at least our customers are sometimes surprised by that when we go to train an AI model to do something. Sometimes the customers maybe have unrealistic expectations about what it could do. We sometimes have to pull back the expectations and basically remind them that if the humans can’t agree, there’s no way the model is going to figure it out.
[00:18:13.800] – Josh
I imagine that comes with the territory. What other kinds of surprises maybe do you imagine engineers are likely to run into as they start experimenting with this technology?
[00:18:25.700] – Eric Cheng
It is very good at doing certain things, but it’s not a magical solution that can do everything. So there are definitely pros and cons. And so I think what we found is we like to blend both traditional computer vision techniques and the machine learning AI-based techniques together. And oftentimes, there are things that are somewhat complementary approaches, basically. So like I said before, what traditional computer vision, some of the advantages of that are that it gives you the ability to directly measure certain quantities or metrics out of what you’re trying to detect, like the size, shape, contrast, things like that. And the ability to do that also gives you lots of different knobs you can turn to directly control the sensitivity of the algorithm and make trade offs and adjust and optimize very directly in an intuitive way. So oftentimes we’ll work with customers and they’ll say, Hey, I want to detect a hole, but I don’t want to alarm if the hole is smaller than this, and I don’t want to alarm if it’s a tear of this shape, something like that, right? That is something that is very easy to directly program with traditional computer vision techniques.
[00:19:49.300] – Eric Cheng
They could just tell us the spec and we can program it in. Whereas if you were trying to do that with a more AI-based approach, it’s a little bit more difficult where you would probably have to collect training data that samples that full space and possibilities in order to really capture the spec that the customer wants.
[00:20:13.100] – Josh
There is a tipping point in there where you reach a level of complexity or simplicity, I guess, could go either way, where you’re able to say, No, we should really just take the traditional approach with this or vice versa.
[00:20:27.820] – Eric Cheng
We often find that the traditional approach will take you maybe 80, 90% of the way there sometimes. We often blend the two approaches where the traditional computer vision approach will do an initial detection. Then once you’re trying to start classifying what that detection is, that’s where you start to add in the AI on top of that. The two work hand in hand and with the computer vision doing some of the initial work to really simplify what the AI model has to actually look at. We find that that works quite well.
6. Business Impact & ROI for Manufacturers
[00:21:05.300] – Josh
Very interesting. I really like that because that’s a topic we keep coming back to is where is harmony between these different, let’s say, generations of technology. What do you think is the best example you can point to for the impact that AI can have on the vision system, both in terms of design and of performance?
[00:21:27.360] – Eric Cheng
It really allows the customer to dream big in terms of where they would want eyes on the production process. What we found is it could be something far upstream in the process where you’re trying to look at the raw materials coming in and detect certain conditions that might contribute to contamination of the raw materials. Those are things that sometimes you couldn’t do before with traditional approaches, and so you can catch defects the few of the better before Then you start to create waste, or ideally before it even gets out the door to a customer, and then you have claims and all that. So it really just expands the range over which you can apply machine vision and allows you to go further upstream and create a lot of efficiencies and prevent claims and waste.
[00:22:23.140] – Josh
Okay. I know you specialize in inspection for building material production. What’s a concrete example that you could give from that space?
[00:22:32.180] – Eric Cheng
Basically, they’re into the production process for several of our customers. There’s basically raw materials coming in that have to get mixed together, blended together in a certain configuration in order to create a certain color, let’s say, a product before that gets applied to the product and sent out the door. We can look basically at that entire process end-to-end and try to catch any thing that is going to contribute to a potential defect in that color blend. That could be looking at the wrong materials, like I said, to make sure that nothing is getting into the raw materials and contaminating them. You could also be looking at a little bit further downstream when the materials get blended together, making sure that the blends are right and maybe there isn’t basically one material that’s run out or isn’t being added in the proportion it should be. Then all the way down to basically the finished product, we can inspect that whole process end-to-end to make sure that the product looks the way it should on the way out the door.
[00:23:53.040] – Josh
Are you able to talk about specialization of algorithms for some of the different verticals that you work in?
[00:24:01.220] – Eric Cheng
We take in very specific problems in the different verticals, let’s say, whether it’s roofing shingles or siding or flooring, and developed some very specialized algorithms to detect very particular defects on those products, whether it’s related to the color appearance or the flatness of flooring tiles or other very specific things. basically, providing these analytics as a service from the cloud has allowed us to customize these algorithms quite a bit rather than needing to ship them something that’s off the shelves that they have to tweak. It’s allowed us to spot things that I think nobody else can spot in these products.
[00:24:52.440] – Josh
We have talked previously with previous guests on this show about that design process and trying to put together a system that’s going to capture all these different types of defects that your end user, your customer is interested in spotting. We’ve talked about the scope that you’re able to capture. Is it 80% of defects? Is it 90% of defects? Is it 95? What are your thoughts on that when it comes to, I’d say, either the level of defect capture that you’re able to achieve using AI machine learning algorithms or the design time that you’re able to save by taking that approach.
[00:25:30.600] – Eric Cheng
Yeah, it goes back to that concept of the last mile that you were talking about earlier. A lot of that last mile work is making trade offs about how you want to tune the system and which defects you want to prioritize because as with most things, the system is not going to be perfect and detect everything you want and nothing you don’t want. So you generally do have to make trade offs. And so that’s really a lot that last mile work that we take care of, which is knowing how to optimize and tune the vision system to go after the highest-priority defects. Oftentimes, we find that a customer may have 10 or 15 different defects that they want to detect, but maybe one or two of those are really the main source of most of the claims or most of the waste or most of the downtime, whatever it might be. We often have to guide the customer towards figuring out what those priorities should be and have them focus on optimizing the system for that. Then oftentimes, once we’ve optimized the system for that, there’s all sorts of other downstream things that the customer has to do now that they can actually automatically detect these things.
[00:26:51.040] – Eric Cheng
So oftentimes they’re not fully prepared to deal with the fact that they’re now detecting the source of the waste. Then they have to figure out, Okay, what do we do with that now and how do we address the process to respond to these detections? And so it’s a very iterative process. And so when you start by just prioritizing the defects in that manner, we can usually go after the ones that will give you the biggest bang for your buck. Then once you’ve stabilized the process in response to being able to detect those, then can just go on down the list.
7. Future of Cloud-Based Vision in Manufacturing
[00:27:32.880] – Josh
Do you feel like… Okay, so it sounds like what you’re suggesting is over the course of time, maybe you’re beginning with traditional vision algorithms as your first-pass design, but then you’re able to iterate on that design. As the customer is ready, you can bring in a different set of tools that chips away at the next 20% or whatever it is. Don’t let me put numbers off. I’m much able to do, but yeah, is that right?
[00:28:03.300] – Eric Cheng
We found it works really well with our cloud-based analytics of the service model is that the production process is always changing, and it’s often changing in response to the introduction of the machine vision system, which is shining a light on where the defects are. Then as those defects start to go away, now you can retune the system to look for the next thing because we’re offering it as a service from the cloud. We’re the ones basically who are working hand in hand with the customer on that iteration process on a continual basis.
[00:28:41.420] – Josh
I think it’s also a good insight in terms of the reality of building these systems and how and when to apply this technology, right?
[00:28:52.440] – Eric Cheng
Yeah, for sure. It’s one thing to have the system be able to tuck on your sleeve and sound an alarm when a particular defect condition is present. But then you have to figure out, Okay, what do I do with that alarm? And so oftentimes we find that even if the system is detecting things very accurately with a low clock alarm rate, there’s still a lot of work to be done to figure out if there is something useful that can be done in real-time or if it has to be a more iterative process over time.
8. Final Thoughts & Where to Learn More
[00:29:24.820] – Josh
That’s fascinating. I’ve really enjoyed this conversation. Where can people e’ll learn more about your work?
[00:29:32.200] – Eric Cheng
Yeah. We have a website, intrinsicsimaging.com. That’s probably the best place to go. You can get in touch with us through there and you can find me on LinkedIn.
[00:29:42.340] – Speaker 3
Right on.
[00:29:43.040] – Josh
Okay. I’ll drop those links in the description for the show. Thank you so much for your time.
[00:29:49.300] – Eric Cheng
All right. Thanks, Josh.
[00:29:52.160] – Josh Eastburn
Okay, we’re going a little long today, so we’re going to wrap it up there. For more from Eric Chang, visit intrinsicsimaging.com.
And until next time, I’m Josh Eastburn for MvProMedia.
