Off the Shelf | The Juice
Zumo Labs presents The Juice, a weekly newsletter focused on computer vision problems (and sometimes just regular problems). Get it while it’s fresh.
Week of June 14–18, 2021
____
Computer vision continues to be a game changer for the retail and grocery sectors. This week Trigo, a startup offering cashierless solutions to existing stores, secured an additional $10 million in funding. Meanwhile, checkoutless system provider Grabango raised a $39 million series B just last week.
These companies typically use object detection and image classification to make their solutions tick. Synthetic data is perfect for training these sorts of models — just ask Amazon, who trains their Amazon Go tech on synthetic data — but the hurdle is generally setting up the initial simulation, especially if you’re not a 3D artist. So, we did it for you. Our own 3D lead, Sammie, introduces our starter shelf simulation in our most recent blog post. It includes everything you need to know to tweak the sim and generate your own synthetic datasets.
____
#SmartRetail
What should you do with all of the Bran Flakes images you just generated using our public sim? Well, if you’re Amazon, you could use them to train the computer vision models powering your prosaically named Just Walk Out cashierless technology. Today is the grand opening of Amazon’s largest Fresh grocery store yet. The Bellevue, WA location is notable not just for its size but also for being (perhaps?) the first “cashierless” store with a cashiered checkout option. That means that customers not interested in scanning their palm for crackers will be able to use something the company is calling “cash.”
Amazon is bringing its Just Walk Out tech to a full-size Fresh store for the first time, via Engadget.
#Deepfakes
In digital photography, “device fingerprints” are subtle but unique patterns imparted by the specific hardware creating an image. The same concept holds in generative models, which means that it should be possible to divine the source of a deepfaked image with enough detective work. Facebook partnered with Michigan State University to develop just such a method of detection and attribution, and this week they published their work. The hope is this technology could help investigate and mitigate coordinated disinformation attempts in the future.
Facebook develops new method to reverse-engineer deepfakes and track their source, via The Verge.
#Deeplakes
If you’ve ever looked at a photograph and thought to yourself, “This would be better if it were subtly moving,” this is the project for you. Researchers at the University of Washington have developed a method for creating convincingly animated photos from a single still. Their process involves making estimations of the past and future of specific pixels and interpolating between them. The model currently works on continuous fluid motions such as water or smoke and will be presented at CVPR this month.
UW researchers can turn a single photo into a video, via University of Washington.
#AutonomousVehicles
After securing another $5 billion line of credit from parent company GM, Cruise has initiated preproduction on their Origin autonomous shuttle. The vehicle is unique among AVs; because it is purpose-built for ride-hailing services, it has no steering wheels or pedals — meaning the regulatory landscape needs to catch up, or it will need an exemption from federal motor vehicle safety standards. Cruise plans to assemble the first 100 vehicles this summer.
Cruise secures $5B credit line to buy electric, autonomous Cruise Origin vehicles from GM, via TechCrunch.
#OpenAI
Large language models like GPT-3 have been called “stochastic parrots,” or, per researcher Yejin Choi, “a mouth without a brain.” Perhaps in response to these critiques, OpenAI aims to improve the model’s occasionally crude behavior by training on a curated dataset that tunes it with more desirable values. OpenAI seems aware that this sounds subjective, stating in a blog post, “Since outlining values for large groups of people risks marginalizing minority voices, we sought to make our process relatively scalable compared to retraining from scratch.”
Improving Language Model Behavior by Training on a Curated Dataset, via OpenAI.
#Eproctor
Imagine taking a final exam in a crowded classroom. Now imagine taking that exam at home, during a pandemic, while proctoring software uses your webcam to monitor and record you for signs of cheating. It has been a rough time for students. So the nonprofit Fight for the Future has launched a website intended to monitor and shame colleges planning to use proctoring software again this Fall. Their goal is to abolish the use of the software, which, beyond being an invasion of privacy, tends to discriminate against marginalized students.
New website tracks which colleges will use eproctoring software this fall, via The Verge.
____
📄 Paper of the Week
GANs N’ Roses: Stable, Controllable, Diverse Image to Image Translation (works for videos too!)
A crazy paper title and GANs applied to anime? Count us in. This paper presents a model that learns to map a content and style code into an anime image. Results show how this method can generate a much more diverse range of styles than other state-of-the-art systems.
A meta-note: we sourced this paper from paperswithcode.com, a departure from our usual arxiv-sanity.com. The team at paperswithcode.com has built something special when it comes to staying up to date with the latest and greatest research, putting more emphasis on benchmarks and code than the paper itself. If you aren’t using paperswithcode.com, you should consider adding it to your literature tool stack.
____
Think The Juice was worth the squeeze? Sign up here to receive it weekly.
