Mirror Mirror | 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 August 2–6, 2021
____
The vice president of an HOA in the Sacramento Valley was a patrol officer for the local police department. His boss, the city’s former police chief, recommended that the association automate their neighborhood watch with license plate readers from Flock, reports Bloomberg.
The tech in question is a cheaper, AI-powered version of similar devices used in traffic enforcement but is publicly (or, rather, privately) available and gives the community’s residents a searchable, 24/7 view of the vehicles on their streets. Local police depend on private communities to install Flock readers and share the data it collects. Flock relies on the police to sell more ($2,500+) devices.
“Essentially an Android phone with a high definition camera strapped to a pole,” it photographs plate numbers that an AI scans for matches against federal, state, or custom ‘hot lists,’ databases of cars that authorities have flagged. Flock has captured photos of more than a billion vehicles in more than 1,200 cities in 40 states.
This week’s roundup reminds us that artificial intelligence is just that — artificial. It can look, but can “see” only to the extent that a mirror can.
____
#BigPicture
A South Korean research team has developed a camera that better creates 3D maps in challenging environments, combining two camera types that individually offer attractive qualities for computer vision. The first is an event-based camera able to capture fast-moving objects, and the second, a fisheye (or “omnidirectional”) camera adept at capturing very wide angles. The team developed a novel approach for the software, poetically dubbed “EOMVS,” that reconstructs the 3D scenes with the event camera instead of rebuilding a scene with images from several angles. Their error rate is just 3%.
This Camera Can “See” the Bigger Picture, via IEEE Spectrum.
#Synthetic Data
Indoor scene recognition has been frustrating programmers and programs for a while now — scene recognition that works successfully in outdoor spaces can’t tolerate the variability of the indoors. For example, some indoor spaces are characterized by global spatial properties (like hallways) and others by their objects (like bookstores). Apple’s machine learning researchers have just released “Hypersim,” a large dataset of photorealistic, comprehensive, interactive synthetic scenes. It fills a much-needed gap in geometric learning and multitask problems, and the generation was half the price of training an advanced NLP model.
Apple’s Machine Learning Team Introduces ‘Hypersim’: A Photorealistic Synthetic Dataset for Holistic Indoor Scene Understanding, via Market Tech Post.
#Mathletes
This year, armchair experts viewing the Olympics are getting real-time stats to bring their analyses to the next level, courtesy of computer vision-powered software from Omega (the Games’ official timekeeper). Omega says it consulted with “scores” of athletes and coaches during development to determine what data was most valuable. Some of the enormous amount of information it tracks: swimmers’ live positions, equestrians’ time in flight, the speed of both the ball and the players in beach volleyball, the complete movement in a gymnast’s pose, and the amount of time in each third of speed climbers’ races.
Numbers Games: More data available to Olympians and NBC viewers than ever before, via USA Today.
#Precog
According to Command leader General Glen VanHerck, the US Northern Command (and eleven other commands concurrently) has completed a series of tests known as the Global Information Dominance Experiments that could give the Pentagon the ability to “predict events days in advance.” The tests combined global sensor networks, AI systems, and cloud computing resources that aim to “achieve information dominance” and “decision-making superiority.” The system — machine-learning-based — can identify changes in raw, real-time data that may indicate trouble, such as a challenged line of communication or an unusual number of cars in a high threat location. The commander says GIDE shifts the Department of Defense’ focus from “defeat” actions to “deter-and-deny.”
Pentagon believes its precognitive AI can predict events ‘days in advance’, via Engadget.
#SwiperNoSwiping
YouTube is banding together in rebuke of porch pirates once more. Ryder Calm Down, a creator who says his content depicts “artificial intelligence to compensate for my own lack thereof,” released a video showing off an anti-package-theft alarm system armed with flour, a truck horn, a sprinkler system, and AI. Ryder’s setup is indeed “simple enough,” consisting of a porch-aimed camera and custom model TensorFlow. But it required a month-long process of teaching an AI to identify “loads of differently sized packages” and to implement a protective whitelist. If his ends inspired you, but the means don’t seem worth the fuss, take a peek at Hugo’s synthetic data-powered package detection experiment (King Arthur not included).
Man defends against package thieves using machine learning AI, flour, and very loud sirens, via PC Gamer.
____
📄 Paper of the Week
Can a Fruit Fly Learn Word Embeddings?
Fruit fly brains are very well understood; they are relatively small (in the order of thousands of neurons) and have been studied for decades. The fruit fly brain typically converts a multi-modal series of inputs (olfactory, temperature, humidity, and visual) into outputs such as motor actions. These researchers re-create-ish this brain structure in code and see if they can get a fly-sized neural net to learn word embeddings (correlations between words and their contexts). And it kinda works! This might seem like a toy problem, but it poses an interesting question: how much should we be looking at biology to understand and design our own artificial learning systems? Now it’s just a waiting game until someone pulls this off with a human brain.
____
Think The Juice was worth the squeeze? Sign up here to receive it weekly.
