The Businesses of Computer Vision

Hashtag #CVPR2022

This blog has given a fair amount of attention to AI research conferences for a while, and with the venues starting to shift back to “in person” crowds there has in parallel been a reversion of the co-located vendor expos to being more commercial in nature. (The online virtual expos in many cases had vendor booths that were mostly hyperlinks and recruitment videos without much in the way of commerce or interactions.) The timing of better attended vendor expos is opportune to this author who finds himself back in the resume distribution game.

IEEE’s CVPR conference (standing for Computer Vision and Pattern Recognition) is a computer vision conference first, so the majority of research and workshops are aligned with that modality, although with recent progress in large scale text to image generative models like OpenAI’s DALLE-2 and Google’s Imagen the modality boundaries in the field are starting to blur. To a lesser extent, the vision modality specialization also extends to the vendor expo.

The following essay is a survey representing the entirety of registered expo exhibitors. These descriptions are offered in good humor and in the interest of improved attendee experience for exposure to a diverse collection of exhibitors. If any of the parties discussed herein who do not manufacture cars are uncomfortable with their portrayal, please feel free to contact the author and he will be happy to strike the entry. If you like reading with a soundtrack there is a New Orleans Jazz Fest compilation link following. Fair warning: there will be jokes. Yeah so without further ado.

Jazz Fest: The New Orleans Jazz and Heritage Festival (compilation, 2019)


Booth: 1006

What is the Metaverse? I mean who knows really. It’s kind of a catch all phrase for the virtual world, where we’ll someday shop, play, socialize, work, all kinds of stuff. But while wearing goggles. I mean that’s one interpretation, there’s others. The metaverse could be a resource for synthetic data to train models for instance. Populating the Metaverse with digital avatars with human-like characteristics is a challenging task given the range of complexity in features and behaviors. We are all unique. (Except for that one Chinese guy that looks just like Elon Musk but that is kind of a freak occurrence, sort of a glitch in the macro meta verse). 3dMD appears to specialize in this task of forming models and representations of human-like diversity in composition and what not. For example they specialize in generating faces, hands, feet, all of those fine grained details that models may want to pay extra attention to, like in healthcare applications for instance. This application requires detailed 3D point clouds and high resolution photography. And apparently they’re good at it. So yeah. Check em out.


Booth: 1229

Data-Centric AI. Kind of a buzz word right now. All of us peons in ML Ops have Andrew Ng and company to thank for organizing a diverse field of publications in the latest NeurIPS workshop, which wasn’t the first and won’t be the last, but happens to be the one where this author was published, so yeah obviously the most important. Anyhoo, Activeloop is all about taming data pipelines for computer vision datasets. Which when you consider that some of them may get into the terabytes is no small thing. Load data, convert data, transform data, interact with images as arrays. But better apparently. Check em out.

Aibee US Corporation

Booth: 1209

Kind of one of those limited info in English on the web circumstances. Aibee appears to be a Chinese startup focused on hot topics like speech recognition, language models, that whole NLP umbrella. If you believe their techcrunch profile, they also happen to be China’s leading AI startup. Oh and they’re backed by Sequoia Ventures. Wow. Check em out.


Booth: 1509

Data-Centric AI. Back in the game. Multi-modal ingest, AI ready outgest (which word I am assuming is the antonym for ingest). Check em out.


Booth: 820

Data Centric AI. You’ve got images, you’ve got video, and you need to figure out how to label partitions thereof. Alegion is here for you. Using ML assistance, human labelers are much much more productive, and your models will perform better as a result. Check em out.


Booth: 1105

Let’s say you are a machine learning researcher who wants to train models around the clock. Eventually with enough utilization you may reach a breakeven point where the cost of owning your own hardware becomes cheaper than training models in the cloud. Of course the tradeoff is that you are now responsible for managing a data science workstation, and all of the drivers and contraptions therein. Which is no small thing. Oh and it will be obsolete in 3–4 years. But just think about all of the good times you will have in the mean time. You’ll never get chilly in the office with your little space heater by your side, and if you need a white noise machine to take a nap those whirling fans have got you covered. AMAX sells model training space heater white noise machines. And they may be just what you’re looking for. Check em out.

Amazon Science

Booth: 827

I think these guys sell books and stuff. Check em out.

Anduril Industries

Booth: 1111

Nothing to joke about, Anduril takes autonomy to the most serious of domains, the defense sector. Air systems. Underwater systems. Where needed. Check em out.


Booth: 1101

Data Centric AI. AI doesn’t work without training data. It is kind of fundamental. What if you don’t have training data? Perhaps Aspen can help. They can source data, they can annotate data, and they’ve completed thousands of projects so they’re probably pretty good at it. Check em out.


Booth: 801

These guys have a pretty good music subscription service. Check em out.


Booth: 1306

Data Centric AI. Your data is your competitive advantage. If you’re not working on one of those NLP models that train on, like, all of the data there ever was, you may be interested in curating. Finding examples and counterexamples. Refining labels and partitions. If you tend to your garden the flowers will grow. Aquarium is there for just that. Curate your data. Inspect your data. Refine your data. Even at the smallest granularities. Because data matters. Check em out.

Argo AI

Booth: 1407

Tired of your rush hour commutes? Argo can help. They are one of the players focused on self driving automobiles. Ready or not here they come. Check em out.


Booth: 1519

Web site down — data collection and annotation. I am assuming they are in stealth mode so will respect their right to privacy.

Borealis AI

Booth: 1020

Financial services. It’s a big field. As far as I can tell Borealis is targeting AI for pretty much all of it, from personal finances to algorithmic trading and a bunch in between. But what is cool is they are not only builders, they also publish. Papers. Open source software. They are looking to contribute to advancing the field. And that’s kind of cool. Check em out.


Booth: 1423

When you think of Bosch what products do you think of? What’s that, you don’t think about Bosch? Well tell you what I’ll try to help. When you think Bosch think industrial. Automotive solutions. Home appliances. Control solutions. Bosch makes things. And Bosch AI makes things that makes things smarter. Yeah so these guys have been exhibiting for a few years now, kind of regulars, good for them they were here early. Check em out.


Booth: 1307

Wait, why did Facebook change their name to Meta again? They were possibly taking the cue from these guys. Perhaps you’ve heard of them by their alter ego: TikTok. That’s right, the Chinese social media video sharing site taking the world by storm. TikTok took AI curation and amped it up to 11. Most dials only go to 10 you say? Yeah but this one goes to 11. And they’re kind of a big deal, like to the point where their platform could someday sway elections. Check em out.


Booth: 918

Data Centric AI. Here we go again. ClearML appears to be honing in on all of those data ops logistics that we know so well. Experiment tracking. Deployment. And a bunch in between. All from a python API and subscription model platform. Don’t believe me? Why don’t you try their free tier and see if it’s worth an upgrade. Check em out.


Booth: 1422

“Hello. Oh I can’t talk right now, can you give me your home phone number and I’ll call you back? What’s that, you don’t want people calling you at home when you’re having dinner? Well now you know how I feel. (Click)” — Jerry Seinfeld. I have to wonder if those recurring cold calls on the Seinfeld sitcom would have gone a little better if they had a platform like Cogito at work to train and monitor their call centers. Because that’s what Cogito does. It is AI powered solutions for call centers. Check em out.


Booth: 1518

Data Centric AI. Cord appears to be automating the annotation of image modality for training computer vision models. Of course there are already sizable startups like Scale that are tackling similar problems, Cord appears to be differentiating by building on lessons learned from the founder’s background in fintech. London based, Y-Combinator alumni, they’re building something that people want. Good for them. Check em out.

Cruise LLC

Booth: 1001

Cruise is somewhat of a prominent player in the field of self-driving cars. Their platform differs from those Tesla guys as they appear to build their models around sensing modalities that include extra equipment like LiDAR for obstacle proximity sensing. Oh and their valuation is pretty solid with backing from several vehicle OEM’s like GM and Honda, so no slouches here. One way or another these cars are going to drive themselves, I hope you’re ready. Check em out.

Datagen Technologies, Inc.

Booth: 813

It is sort of a distinct convention of supervised learning to use synthetic data for training. I don’t know if you’ve played X-Box lately, but those graphics engines are getting pretty good. And companies like Unity and NVIDIA are building whole libraries of pre-packaged virtual mirrors of real world objects. The aisles in your supermarket. Medical imagery. Driving down the freeway. If it is an environmental envelope with known boundaries and conditions it will probably be possible to build some form of synthetic mirror to train or augment training of applications like supervised or reinforcement learning. And Datagen is building a platform for just that: building synthetic datasets for computer vision. Check em out.

Datatang Technology Inc.

Booth: 1207

Ever heard of Mechanical Turk? It is an Amazon platform for anonymous contract workers to rent their time by the hour to perform menial tasks. Do you really want to trust your data annotations to the anonymous? Companies like Datatang are here to step up with pre-labeled data or thousands of specialists from around the world to build from scratch. Data annotation and labeling. And they are specialists so you can expect they are using semi-automatic tools for improved efficiency. Check em out.

Booth: 909

Do you need some artificial intelligence? Well of course you do but it’s not like they have an artificial intelligence store where you can just go shopping. Wait what’s that? is an AI marketplace? Well perhaps they do. Go shopping for data. And models. Oh and honey while you’re at the store can you pick up a gallon of milk? Check em out.


Booth: 1227

There is an open question of what mainstream self driving car platforms are going to look like, and one of the biggest questions remain what are the sensing platforms required. Us humans primarily drive with our eye cameras, and for the most part do pretty good outside of like heavy rain and stuff. LiDAR sensors take it up a notch by feeding physical 3D line of vision models, with biggest tradeoff being the cost of the hardware. Dielmo appears to specialize in all of the considerations around packaging LiDAR systems. Software. Development. Solutions. Check em out.

Digital Divide Data

Booth: 1205

I tell ya this is a catchy company name. The digital divide is a real thing, after all the future is already here, it just isn’t evenly distributed. How do we bridge it? DDD is looking to bring the future to the underserved with content, data, and research — and perhaps raise a few people out of poverty in the process. Check em out.

Disney Research

Booth: 822

These guys have been building the future ever since it was the past, just check out EPCOT if you want to see what I mean. The mouse has the most recognizable ears since commander Spock, may he live long and prosper. Check em out.

Exxact Corporation

Booth: 907

If you look at the history of computing, I mean like the entirety of it, the big paradigms have centered around distribution and scale of infrastructure. The mainframe multiplied to the PC’s. Now we’re back to data centers. These Exxact guys appear to be looking to cover all of their bases in their product line. Need a workstation? Two? An entire AI Cluster? Well perhaps they can help. Check em out.

Flex Logix

Booth: 1503

With the advent of deep learning in the 2010’s so much of the hardware stack became centered around NVIDIA’s predominant GPU hardware and CUDA software stack. Everyone wants to carve out a piece of it, and Flex Logix’s angle appears to be centered around hardware for running inference. Edge computing devices have special needs for efficiency, throughput, and heck if you want to be competitive cost too. Want to run some inference? Check em out.


Booth: 1026 , 1800

Furiosa is in the chip design business. And the vision model performance business. They appear to be targeting throughput / latency performance on common vision tasks like classification and object detection. All with their own hardware and software stack. Keep your eye on this space. Check em out.

Glass Imaging Inc.

Booth: 713

It is easy to take for granted the amazing pace of portable camera progress that has been made with smart phones. I look at pictures I took five years ago verses those I took now and it is night and day. Especially at night. These Glass guys appear to be at the forefront, seeking to take those highest end “SLR” cameras and miniaturize so that they fit in your pocket. Before you know it we’ll all be taking pictures of black holes from our back yard. Check em out.


Booth: 721

These guys are bringing the Dewy Decimal System to the twenty-first century with artificial intelligence, and they’re doing it with an empowered and independent workforce — most of whom have a deep mind of their own. Check em out.


Booth: 1010

If you want to see what a macro micro chip looks like (play of words, kind of like jumbo shrimp, yes this author is very funny, sorry I digress) you need to see Graphcore’s “IPU” chip. Seriously it is huge. They’re accelerating model training at the largest scales. After all, kind of like what they say in Texas, everything is bigger at Graphcore. Check em out.

IEEE Computer Society

Booth: 1330

IEEE started as a society for electrical engineers nearly 50 years ago, and have since grown to become a hugely influential standards board and professional society found in related domains, organizing global conferences across industry. Now if only someone would put together a research conference dedicated to computer vision. Oh wait they already did. It is this one. And we all have IEEE Computer Society to thank for it. Thank you IEEE CVPR organizers!

Imatest LLC

Booth: 921

As computer vision becomes formalized and industrialized, we can now all enjoy formal specifications, quality control measures, and other trappings of industry. I mean if we’re going to trust self driving cars with our families as passengers for instance one would hope we could thoroughly and professionally validate the hardware. Imatest appears to be looking to do just that. Image systems testing and validations. Because quality matters. Check em out.


Booth: 807

Data Centric AI. iMerit was one of the more extensively verbose websites I’ve come across, but the gist of it is that they offer annotation solution suitable for mission critical applications, with expert in the loop, scalability, you know all those good things. And I think they pay their web developers by the word. Check em out.


Booth: 919

It is not a typo that the V in iniVatino is capitalized, in fact it is a quite clever allusion to the Vision domain for which iniVation specializes (kind of like how every quantum computing startup has a letter Q in their name somewhere, just because). iniVation’s product line appears to include some specialized applications like eye tracking (I like to Zoom on a separate monitor than the one with the laptop camera to avoid this type of intrusion). They also have precision “laser” scanning for 3D imaging (and yes, as is a custom adopted from Dr Evil, one should not use the word “laser” without including air quotes). Yeah so that and other custom vision solutions at your fingertips. Check em out.

Inspur Electronic Information Industry Co., Ltd

Booth: 1413

We’ve already seen hardware providers offering local data science workstation solutions. Inspur appears to be focussing on the upper end of the spectrum, with servers and rack scale systems. They talk about custom solutions for integrating data center hardware and cloud services. And all of this from their Chinese headquarters. As models continue to scale into the overparameterized regime, someday these rack mounted options may be the minimum for those that opt to own their own training equipment. Check em out.

Intuitive Surgical

Booth: 1211

There is no more delicate application for the miracle of the human hand than those procedures performed by life saving surgeons in the OR. How could we expect a clumsy robot to act in this place? Intuitive is seeking to do just that with their Da Vinci line of robotic surgery assistants. Why would you want a robot surgeon? Think precision, repeatability, and they never get tired. Just make sure you don’t trip over the chord on your way to the keyboard. Check em out.

Kinetic Vision

Booth: 1008

I know with a name like Kinetic Vision you may think these are the kings of the convolutional net, possibly recruiting Yann LeCun to take over as head of research. It appears there is actually more going on here. Kinetic are product developers. What does that mean? They are here to help with the entire product development journey, from AI to packaging to IP, if you have an idea, they have the team to get you to market. Check em out.


Booth: 1522

While most of us like to think about AI as software without the need for software engineering, that reality is still a distant future. Even in the few shot learning paradigm, integrating AI solutions into real world systems requires packaging and pipelines to make it robust. Kitware is here for just that, they can help you design and develop AI integrated solutions, and they have an expansive portfolio of open source solutions to build on to make even complex projects feasible. With applications from medical to energy to government, these guys are the real deal. Check em out.


Booth: 1523

It is hard to understate how critical an industry has become chip manufacturing, with nearly every other part of the economy reliant in some fashion. KLA appears to be pioneers at the frontiers of silicone process management — reminiscent of the old days when “Silicon Valley” was a moniker that meant just what it said. As circuits get ever more nano, AI is here to help with everything from chip design to quality control, and KLA is here to help with getting help from AI. Check em out.

Kodiak Robotics Inc

Booth: 1318

It remains to be seen which parts of the transportation economy will be first to fully automate. Given that their autopilot has been around for several decades, one would think that airplanes could be good candidates, until you consider that landing is a little trickier than flying and oh yeah there are 100+ people on board. The Uber taxi type application of ride hailing apps is a little more scalable, but unfortunately they’ve got all of that city driving to deal with, which for self driving cars is kind of the hardest part. Other than railroads, a big application will likely be interstate trucking, after all freeway driving is a problem mostly solved. The tricky part is that loaded semi’s are much less forgiving in outlier scenarios. Fortunately Kodiak is here to help, now if we could just figure out how to solve the last mile challenge. Check em out.


Booth: 1310

Data Centric AI. You’ve got all this data, heck some of it is even labeled. Let’s say you want to label the rest. That means you need to manage access, and you need to manage processes, and you need to do it all fast. People may not realize this, the amount of manpower that may be thrown at a labelling application is enormous. And costly. And getting it right is impactful to model performance. Data labelling can be a thankless job, in the worst cases resembling the digital equivalent of virtual sweatshops to their workers. The better the tooling we can provider to these workers, the better the outcome on many fronts. Let’s get it right. Check em out.


Booth: 1119

Life is good as they say. At least it is for those consumers that may have a LG big screen TV in their living room. That is a part of it, but LG is more than just televisions. They build appliances, they build air conditioning equipment, they are basically building the suburban dream. Oh and apparently they’re now AI powered to boot. (I am reminded of Disney’s Carousel of Progress for some reason). When you think of Korean consumer tech LG is a big deal. Check em out.

LUCID Vision Labs, Inc.

Booth: 1022

Smile, you’re on camera! At least you will be if LUCID has anything to say about it. Offering a range of high performance sensors for self driving cars and many other industrial vision applications, they don’t stop there. They help with software development too. Now say cheese. Check em out.


Booth: 1304

If you’re like me you might be a little confused about the difference between Mathematica and MatLab. Well I think Mathematica is a little more symbolic manipulations and explorations inclined, whereas MatLab is more centered on like infrastructure requiring mathematic operations or high precision experiments. I mean to be honest I’m also a little fuzzy about where the Wolfram Language leaves off and Mathematica begins but I digress. We’re here to talk about Mathworks, which is robust infrastructure for all of your mathematical needs. And if you spent any time in grad school at all, you know how to use it. What’s that, you learned something useful in grad school? How about that. Check em out.


Booth: 1019

I mean all of that PyTorch and social media infrastructure aside, these guys are basically building out the metaverse, which if I learned anything from the Spiderman movies means we’ll soon be meeting a few of our doppelgängers who each have haircuts of a varying level of success. Oh wait sorry that’s the multiverse I was thinking of. Check em out.


Booth: 911

When you think of semiconductors it is tempting to focus on those sexy microprocessors and GPU’s that make things go whir. But where would the von Neuman architecture be without memory? A computer with memory but no processor is still Turing complete (that’s what a Turing machine is after all), one can’t say the same for a circuit with a processor but not memory. These Micron folks are here to ensure that we all have enough memory to remember this fact, and if $150B in manufacturing infrastructure in the pipeline doesn’t convince you I don’t know what will. Check em out.


Booth: 1313

These guys have the longest running software product in the history of computing, still going strong. Oh and they’re on the cloud too, I azure you. Check em out.


Booth: 1018

We’re all aware of the logistics magicians at Amazon and how they have countless robots in their warehouses and distribution centers picking and carrying boxes and racks in a high traffic orchestrated dance of automation. Motion2AI is here to bring that capability to the rest of us. A warehouse to a human can appear as a chaotic maze, with hidden views and corridors into the endless, perhaps carrying Noah’s lost arc down one abandoned turn. To the computer it is just a database, and their colleagues another appendix in the collective organism. Don’t believe me? Check em out.


Booth: 727

Self driving cars are coming, and whether they primarily manifest as consumer owned appliances or fleets of taxi infrastructure remains an open question. We all have Uber to thank for trodding the difficult path of pushing back against unionized taxi regulations city by city, and their thousands of contracted drivers to thank for setting the stage for their own eventual obsolescence. Fortunately for supply and demand, just because Uber set the stage, that does not necessarily mean they will have a monopoly on ride hailing. Motional is developing an autonomous vehicle for the taxi market, now if you excuse me I need to hail a ride to the airport. Check em out.

MUKH Technologies, Inc.

Booth: 1326

Ok the reality is that AI is going to have applications that are trivial or entertaining, and then there are going to be more serious types of applications. Law enforcement, surveillance, and other security mechanisms. Facial recognition software has proven to be controversial at times, especially owing to potential screening bias, but in high traffic public facilities it is going to be an invaluable tool for security whether you like it or not. The real question is where do we draw line, do we want to allow facial recognition at the grocery store too? How about on the street corner? The point is that facial recognition is already here, and the extent of surveillance that we want to normalize is probably worth some debate. In the meantime, MUKH has a packaged solution that works off the shelf. Check em out.


Booth: 1529

We in the US hear a lot about Chinese doppelgängers of American industries, at least to this author perhaps less so about Korean equivalents. NAVER is an internet conglomerate from the Korean market. If you can think of an American web portal, they likely have something similar. Search engine, messenger app, social media, you know — the internet and stuff. Now the real question is whether they have better bar-b-q. I mean Texas makes some pretty good baby back ribs, I honestly don’t know how NAVER could compete with that. Even so, check em out.

Neural Magic

Booth: 1223

This is kind of a fundamental point, supervised learning has a natural boundary in the scope of model training and inference. In many cases, a model may be trained on one computer and have inference run somewhere else entirely — say an edge device, a local server with access to sensitive information, there are all kinds of scenarios. The point, if I have one, is that the paradigm of model training on GPU hardware has naturally defaulted to inference on GPU hardware, even though the computational intensity of training and inference can be vastly different. Neural Magic offers solutions to “sparsify” trained models in order to open the door for even large models to have inference run on CPU processors — which are generally more affordable, prevalent on edge devises, and you know, just kind of a first class citizen in consumer PC’s for instance. CPU inference opens the door to all kinds of distributed applications, and Neural Magic is making it a reality. Check em out.

Openedges Technology, Inc.

Booth: 922

Trying to be snarky and kind of struggling how to do so with such a pure tech play. So am just going to put on my technologist hat for a second. Openedges is developing IP for edge computing applications, more particularly, DRAM optimized neural processing units and associated software. I mean how does one make a joke about that? Oh well, check em out.


Booth: 1419

You know I never bothered to look up what the acronym LiDAR stands for. I’m assuming R stands for radar, right? Like it drives automobiles radar. That sounds about right. Cool well glad we figured that out. It is what Ouster does, they build LiDAR sensors, and based on their website it looks like they have several options for different market segments, with both low and high volume sourcing options. They sell really good like it drives automobiles radar. Check em out. *editor’s note: “light detection and ranging”

Panasonic AI

Booth: 806

Panasonic AI appears to be an all purpose research lab supporting Panasonic, a Japanese industrial conglomerate manufacturing all kinds of consumer solutions like appliances, automotive, B2B, the whole nine yards. They like to build stuff, and with a dedicated AI research team you can be assured that they are doing it at the cutting edge of technology. Fun fact, I used to have a Panasonic stereo amplifier — it was awesome. Panasonic is the real deal, check em out.

Parallel Domain

Booth: 1301

One of the hardest parts of training models for self-driving cars is that you’ve got so so much data, millions of hours of driving, and yet 99.99% of it constitutes a vehicle traveling 5 mph over the speed limit and following three lengths behind the car in front down a single lane along the interstate. Or it’s just bumper to bumper stop and go. Or it is the last mile of slow driving through a neighborhood. And a model could be trained on this data ad nauseam and learn nothing more than how to follow three lengths from the car in front. Automobile safety is all about edge cases. When the flashing lights are behind you. When a bicyclist is on the median. When a pedestrian is jay walking. When the unexpected happens. Parallel Domain offers a generative pipeline that can be tailored to custom sensor arrangements for generating all of the synthetic training data you might want for these edge case. Do you want level 4+ autonomy in your fleet? First your model is going to need to expect the unexpected. Check em out.

Perceptive Automata

Booth: 719

Edge cases for training self driving cars come in several flavors. You’ve got the random strange object in the roadway, you’ve got weather conditions, or even other cars not following rules of the road. Each of these are challenging, but they pale in comparison to the most intractable actor of them all: the quixotic pedestrian. They may be exercising, they may be jay walking, or heck they may even just be teenagers. The point is if a car is going to accommodate those people sharing the road, it will need to find a way to predict their behavior. I know what you’re thinking: good luck with that. The Perceptive Automota solution attempts to make use of all available signals at hand to form a theory of theory of mind so to speak. These signals could be body posture, leg movement, focus of attention, all those things that us human drivers take for granted as coherent. It’s the only way. Check em out.


Booth: 1321

It’s always easy to anthropomorphize machine learning and consider its capabilities as somehow approaching or resembling those of people. The reality is that the machines will have capabilities of input and analysis extending well outside the range of what we experience as people. Consider the range of the visual wavelength spectrum or the micro spans of time perceptible to the human mind. Computer vision will be able to operate well outside of this envelope, especially with the high performance camera sensors available from Photron. Check em out.


Booth: 823

What is Pinterest? You know I’m not sure. It is kind of like social media, it’s kind of like a shopping, music, and book recommendation engine. I mean really it is a platform that serves all kinds of use cases. And the only realistic way to serve so many diverse use cases to a large user base is to do it with machine learning. (If you like this observation feel free to pin it to your board.) Check em out.


Booth: 818

This is what I like to see. Plask appears to be an accomplished startup in the truest form. With a focused product built on AI algorithms, refined through iterations to master a distinct niche. The chasm ahead just begging to be crossed. And what do they do? They capture video and translate human poses to animation. I know it sounds a little less grandiose than some of the other exhibitors here, but you know what this is how sustainable businesses come about. Check em out.


Booth: 1323

If you’ve ever seen Jurassic Park, you may be familiar with the convention that when a T-Rex is approaching the safest tactic is to stay in a completely still crouch, even as he leans forward and exhales a smokey breath. A T-Rex’s vision is movement based after all (which I’m assuming we know from fossilized bones somehow?). Yeah it is a good movie we’ll allow some artistic license. Anyway Prophesee appears to be offering a digital t-rex camera, which they call the neurotrophic chip, which saves energy and overcomes limits of conventional convolutional networks by having a sensor mechanism that activates for events instead of images. Just like a T-Rex. Next thing you know they’ll have a velociraptor chip that pauses when you hold your hand out and say like “easy there” or something. Check em out.


Booth: 1401

The march of progress in cellular signals has been relentless. 3G. 4G. LTE. 5G. What’s coming next? I am guessing it will either be 6G or like LTE-2 or something. I bet the Qualcomm folks know the real answer. After all they’re the ones building the communications chips to do so. Check em out.

Retrocausal AI

Booth: 1308

The ability for machines to collaborate with humans in industrial settings like a factory floor share common challenges when the process envelope of conditions is coupled with unpredictability of outlier inputs or human interactions. Retrocausal AI offers solutions to model a process envelope, which allows for intelligent feedback signals to their human collaborators, which could benefit training, safety, all kinds of stuff. Think of them as just another kind of hardhat, gloves, safety glasses, and steel toed boots — but one that thinks. Check em out.

RoboSense/Suteng Innovation Technology Co., Ltd.

Booth: 1319

It is easy to dismiss LiDAR systems as just another piece of hardware, sort of a commodity in the context of the self-driving vehicle application. The reality is that performance variances can have meaningful input on resulting safety. Would you rather be a passenger in a car that can sense 100 meters down the road or 1,000? How about the difference between captured images every quarter of a second verses thousands of a second (ok full disclosure I am not a domain expert, these specific numbers are kind of imaginary to demonstrate a point). Further, it’s not just the hardware of the LiDAR that matters, it is also how it interacts with and is operated by system software and chipsets. It requires a cohesive system design to operate near state of the art. The friendly folks at RoboSense are here to help. Check em out.


Booth: 1013

Intermediation. It is kind of a MBA term. Think of it in terms of you have a service that you could either contract with directly, or a middle man comes in who is more convenient and oh by they way they also have a generic version of your product in their catalog. This is just an example. Roku is basically trying to aggregate all streaming services — Netflix, Prime, Disney, you name it — and rebundle the great unbundling, with the help of a USB stick and a comfy remote. Yeah so where does AI come into play? Well if they’re going to sell advertising for their generic versions they need to understand your tastes. Movie recommendations, show recommendations, that kind of stuff. And oh yeah, they’re offering this marketplace on low cost hardware without added cost to the consumer. I think it’s a great business model personally. I’m not sure what it has to do with computer vision, but you know, everyone is welcome. Check em out.

Saudi Federation for Cyber Security, Programming & Drones

Booth: 1700 , 819

This author’s jokes notwithstanding, CVPR is a very serious conference for next generation research and important applications, with attendees ranging from startups to sophisticated actors in domains like cyber security. I am guessing this Saudi Federation has state of the art applications and talent, and if you want to do business with them this conference is a window of opportunity. Check em out.

Scale AI

Booth: 901

Data Centric AI. There are a lot of exhibitors focused on data labeling and annotation infrastructure. Scale was one of the first, and remains one of the most successful. Seriously, they’re like a unicorn startup last I checked. No joke. Ok just one joke. Wouldn’t it be cool if when you’re trying to lose weight your scale had some AI that could kind of fib a little bit about your current weight to motivate you a little more. I’m just saying that could be a good product line extension for Scale AI. Think about it guys. Check em out.

Booth: 1505

Data Centric AI. What’s that, you have a data set management and labeling application but you prefer a python API? If only someone could, oh wait here we go, Segments’ are just what you’re looking for. Build better data sets. Do it in a robustly defined API. Start your trial today. This message brought to you by Check em out.

Snap Inc.

Booth: 1322

As we migrate more and more of our personal lives into the meta, there is a certain reality we need to come to grips with. People are inherently foolish. They look foolish. They act foolish. And you know what, being foolish is a big part of what it means to be human. It is a good thing. Not to worry, SnapChat is here to serve you slices of humanity in a short term setting, for impermanent memories and goofing off with your friends. Have a good time without regrets. After all as they say, if it happened in Snapchat it stays in Snapchat. Especially if it is a picture of something that happened in Vegas. Check em out.


Booth: 1201

Besides inventing the original iPod (known then as the Walkman), Sony is a lot of things to a lot of people. If you’ve been lucky enough to get your hands on a Playstation 5 you’re probably not reading this since you’re addicted to video games by now. Or if you’ve been to the movies lately you may have seen the original Spiderman franchise. But wait, there’s more. Electric cars. Robotics. Televisions. Computing. They are a conglomerate in the purest sense of the word. And you can be assured they are bringing artificial intelligence to bear across the entire franchise. Check em out.


Booth: 1520

Springer is a mega corporation conglomerate selling books by the millions and with warehouses around the world. They also have a cloud computing / web services arm and are transforming retail. Oh wait, sorry, that’s Amazon I was thinking of. Springer is in a different segment of the book publishing business. The publishing segment. Yep, journals, textbooks, if you want to learn about machine learning they can help. If you want publish your phd thesis, they can probably help with that too. Check em out.


Booth: 1328

With generative models now so capable, and with the reality of deep fakes on the horizon, Steg is addressing an important question: how can we trust our own eyes anymore? One solution involves a formal provenance verification by embedded meta data. Don’t believe me? A Steg certification can be encrypted in the pixels, hidden in plain sight as they say. Check em out.


Booth: 808

When most consumers think about self driving cars, they probably just imagine a simple input of video and perhaps LiDAR in, and output of steering wheel and gas pedal position. In practice, the application has all kinds of tiers of intermediate data processing and inferences even before considering which way to steer or how fast to go. Lane detection. Traffic sign and stop light interpretation. Adjacent car trajectory projections. Pedestrian evolution evaluation. And STRADVISION is working to modularize each one of these applications for a plug and play solution. You know, like legos. Check em out.


Booth: 1300

Data Centric AI. Wow that sure does come up a lot. Yes you got it SuperAnnotate is a turnkey solution for all of your data management needs across the entire lifecycle, from curation to annotation to integration and everything in between. What sets this one apart from the others? I don’t know. But they can handle any data format and have an enterprise ready security first platform. Sounds useful. Check em out.

Superb AI, Inc

Booth: 1109

Data Centric AI. Here we go again. Wait don’t know if this is new, they talk about automation. As in automating computer vision data pipelines. So perhaps that’s unique? I bet their sales department could tell you more. Check em out.


Booth: 1521

What could be more fundamental to a functioning society than a well-oiled agriculture machine? Syngenta in an agriculture first firm, offering seeds, herbicides, basically everything a midwestern farmer could want. Wait, what does that have to do with computer vision? Well until recently absolutely nothing. However consider how much herbicide could be saved from our soil if the application was targeted to blossoming sprouts. That’s right think robot farmers, looking through their camera vision and spotting the weeds and pests, targeted for curation. A more sustainable future, it is coming and Syngenta could be the ones bringing it to you. Or so I assume. Check em out.


Booth: 1501

If you are looking for a pure-play domain expert in computer vision, the researchers at Synthetaic appear to the the real deal. With a dedicated focus to image classification, Synthetaic is building solutions where even non-technical teams can run applications. Because not everyone speaks the language of Python. Would you rather spend 5 years in a phd program or hire Synthetaic to run a few lines of code? That’s what I thought. Synthetaic: bringing computer vision to the masses. Check em out.

Tangram Vision

Booth: 1421

Computer vision at a high level can be segregated between boundaries of neural network evaluation as fed by some sensor’s input. In real world practice, the channeling of sensor input into the network is where most of the complexity comes into play. Every sensor needs a driver and an I/O method and oh yeah you don’t get the benefit of tutorials and textbooks to help, this is a micro application with a small addressable market so big players aren’t much help. Fortunately there’s resources like Tangram Vision to provide the glue between your sensor’s logistical handling and the neural network itself. Sound useful? That’s because it is. Check em out.

Teledyne FLIR LLC

Booth: 1418

The paradigms of convolutional learning started out in the image modality, but in industrial practice are now extended to all kinds of applications for sensory evaluation, including both static and time series inputs. Teledyne does not appear to limit focus to a particular type of sensor or application, in fact they appear to be active across a broad band from healthcare to industry. And they’re doing it with off the shelf or custom engineered sensory solutions, and they’re doing it at scale. After all every spotlight we shine on the playing field improves an AI’s ability to score a goal. Perhaps you should pass Teledyne the ball. Check em out.

TELUS International AI Data Solutions

Booth: 1420

Data Centric AI. Let’s say you’ve got training data and need to annotate labels. Easy so many options to choose from. Now let’s say you’ve got data at scale, say terabytes of video footage. How does one go about annotating such scale. TELUS has a global community of over 1 million contracted annotators. That is not a typo. And they can work across a wide range of languages and dialects. Someday we’ll have computers as smart as the humans, in the meantime building out applications requires a human in the loop. Or in the case of TELUS, a lot of humans in the loop. Check em out.

Tesla, Inc.

Booth: 1027

The cool thing about watching Elon Musk interviews, besides the occasional sarcastic quip, is that he is beyond transparent of operational details. When asked seemingly any question he shares details of design and bases of decisions, and even long term plans and objectives. It is as if he is speaking to his investors and workforce not just through inter company memos but directly through his media presence. This dates all the way back to the infamous master plan memos drawing out the intended product line progression from the roadster to model 3, all of which came to fruition just as planned. Projecting the specific timing of realized full self driving autonomy is about as easy as projecting the timing of AGI, but if there is anyone that has a sufficient financial incentive to get it right, it is Tesla. Incentives matter, and with the stakes at play for Tesla, who has been selling self driving autonomy even before it is ready, they have all of the skin in the game in the world to ensure they are motivated to get to that threshold. Check em out.

Toyota Research Institute

Booth: 715

As the current largest car manufacturer in the world, Toyota finds themselves in an interesting position with so much industry change on the horizon. In the last year most of their competition has announced some form of planned product line progression from gas to electric drive trains. User interfaces are rapidly evolving towards the central display and OS interface as opposed to the modular stereos and scattered buttons of not too long ago. No one knows exactly which self driving platform will win the horse race to functionality, and most of the industry is setting the integration of sensory infrastructure on hold as a result. Within this context, they also just rolled out a new subcompact SUV which is actually kind of nice for the price point (that is assuming the Glendale, CA local sales team will honor the advertised price and not nickel and dime you even after a handshake just because they can see that the purchaser is 8 months pregnant and her older brother flew all the way from Florida to help her get a car that she wanted to get the baby home from the hospital, which by my experience is not a given). If you are interested in self driving vehicles I expect Toyota will have a big part to play in what’s to come. Check em out.


Booth: 1511

Self driving long haul truckers have a lot of benefits to offer verses a manned vehicle fleet. Besides the obvious talking points — the computer doesn’t get sleepy or send text messages — they also just perform better. They stay in the center of the lane. They apply a more consistent throttle, and as a result they achieve material fuel economy savings, especially at lower speeds. And they do all of this without having to worry about labor shortages of trained and licensed drivers. TuSimple is smart and keeping it simple. Check em out.

Unity Technologies

Booth: 812

Unity Technologies has one of those really interesting positions where they are straddling between multiple applications of a common platform. Whether for advanced performance video games or synthetic training data, their graphics and physics engines are here to present a realistic synthesis of the real world, and serve that model in a digital graphics environment. You know NVIDIA’s GPUs kind of started out that way, selling as video game hardware then repurposed for training models. If you want to build a metaverse, you need a model, and if you need a model, you need a graphics engine. I expect Unity will be around for a long time to come. Check em out.


Booth: 701

Data Centric AI. It’s kind of a popular theme. V7 has a catchy name, and a polished platform for labeling data. Smooth performance, low lag, visual data serving for annotations and labeling. Just what you want in an annotation platform. Check em out.


Booth: 1513

When you think of France what industries come to mind? Ok besides good cheese and baguettes. How about automotive? I mean they don’t exactly have the Lamborghini and Ferrari brands to show off, but Valeo is a big player in automotive manufacturing, with all kinds of products built for manufacture or aftermarket. You know, the nuts and bolts of things figuratively speaking. And they will not be on the sidelines as self driving comes to the forefront. The Valeo research center is in the heart of Paris, where writers go to write and painters go to paint. And coming soon, where cars go to drive themselves. Check em out.

Vision Components GmbH

Booth: 923

There are a few ways to look at industrial sensor supply chains. The cameras could be sourced as a fully packaged system ready to be bolted in place. Vision Components appears to be more interested in sensor subsystems, like cameras and chipsets without the surrounding packaging. You know, for a purchaser to embed in their own system design. Just another way of doing things. Check em out.


Booth: 1219

Data Centric AI. Do you want to run training experiments? It will help to target interventions if you better understand the nature of the dataset itself. Voxel helps you visualize data set groupings and scenarios so that you may understand the nature of what it is your model is trying to accomplish. Oh and they also are active contributors to the research community by way of managing benchmark data sets like ActivityNet and supporting the Coco dataset. They support us, we should support them, as one good turn deserves another. Do I have your support? Check em out.

Weights & Biases

Booth: 1507

Data Centric AI. The practice of data science is one with no single settled form. In general, key challenges include data set preparation, model architecture and parameter optimization, and then rolling out in production. Weights & Biases are an ML Ops platform that appears to specialize in the model development scope. Run training experiments in an organized and accessible manner, with an interchangeable python or graphic API. And this author is comfortable sharing that he has seen them around for a while and they appear to know what they’re doing. When you’re ready to graduate from logging training experiments in some massive spreadsheet where you don’t even know where to begin when looking at it again a few months later, perhaps you can give Weights and Biases a try. Check em out.

Woven Planet

Booth: 707

What’s that Toyota? You weren’t satisfied with my first writeup so you formed an entirely separate research group just so I would have to write about Toyota again. Geez some people are so needy. Ok if I must I must. Just because an automobile manufacturer’s primary business is transportation does not mean they should necessarily limit their research to vehicle automation. Every aspect of industry, from white collar to manufacturing, is eventually going to be impacted by the AI revolution. Toyota is smart to channel research in a wide band beyond basic driving concerns. And Woven Planet is a vehicle for just that purpose. Check em out.


Booth: 1327

The Zoox self driving platform is a wild design. Seriously it is a ground up reworking of what it means to transport people. No drivers seat, no passenger seat, just a pair of inward facing arrangements for conversation instead of staring at your phone the whole drive. Isn’t that cool? Wait a second, if there are two directions for seating, where do we put the TV? On long road trips are we just supposed to stare at each other like a zoom call or something? I can barely hold a conversation for half an hour, let alone the two hours from Orlando to Gainesville for the annual Orange and Blue game. How is this going to work? I’ve got it, let’s put a mirror divider right in the middle and can just use those two hours to look inward and contemplate about what an obnoxious blog post this turned into. Check em out.

Editor’s note: If there is anyone that the author did not manage to offend with this essay, please leave a comment below, and he shall taunt you a second time.

For further readings please check out the Table of Contents, Book Recommendations, and Music Recommendations. For more on Automunge:



Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store