Olympic Metal| The Juice

____

Did you hear about the half-court shot that someone drained during the US v. France men’s basketball game the other day? It might not seem newsworthy, given that the Olympics brings together humanity’s most talented living athletes. But the athlete in question was neither living nor human. Toyota’s Cue 4, a cyclopic, burnt Belgian waffle of a robot with nightmarishly huge hands, took to the court during half-time to remind the world that even our gold medal champions won’t be enough to stop him and his ilk.

Incidentally, the Olympics has long been a testbed for this sort of advanced technology. The first photo-finish camera, the “Magic Eye,” was introduced at the 1948 London games. It was an effort between the British Race Finish Recording Company and Omega Timing. More recently, Omega has been working on computer vision technology to track beach volleyball, which will officially launch at this year’s games. They boast a 99% accuracy but admit that occlusion tripped them up during training.

How will the system perform now that it will be tested on an international crowd of different races and genders? Time will tell.

____

USA Surfing brought home an impressive victory after the sport’s Olympic debut this year. It was, of course, the result of an immense amount of training and hard work — but also something else. The team’s medical director told the WSJ they’ve been searching hard for solutions to improve athlete performance while mitigating uncontrollable variables like wind and tides, and that “taking a deeper dive into analytics and data was our roadmap.” When biomechanics revealed how critical imbalance is, they began developing a machine learning system that assesses maneuvers and stance. It seems it has paid off.

How Olympic Surfing Is Trying to Ride the Machine Learning Wave, via The Wall Street Journal.

Cassie, a very efficient multimillion-dollar robot with “knees that bend like an ostrich’s,” has just made history. It completed a 5K at Oregon State in just over 53 minutes untethered and on a single battery life. Cassie is powered by a combination of biomechanics, existing robot control approaches, and new machine learning tools, and is the first bipedal robot to use ML to “control a running gait on outdoor terrain,” according to OSU’s Dynamic Robotics Laboratory. Jonathan Hurst, a robotics professor who directed Cassie’s development with a grant from the U.S. Department of Defense, says the holistic approach is “incredibly exciting” and will enable “animal-like levels of performance.”​

Bipedal robot developed at Oregon State makes history by learning to run, completing 5K, via Oregon State University.

Google AI announced on Wednesday the launch of an open-source dataset that contains the locations and footprints of more than 500 million buildings across Africa. Having an accurate record of this type is vital for disaster and medical support, census information, and environmental research efforts, but informal settlements, rural areas, and small or indistinct structures can make detecting buildings with computer vision challenging. So, Google AI manually labeled 1.75 million buildings in 100,000 images and trained a model with a U-Net-style detection pipeline that is “bottom-up…first by classifying each pixel as building or non-building, and then grouping these pixels together into individual instances.” The project was led by Google Research, Ghana as part of AI for Social Good.

Mapping Africa’s Buildings with Satellite Imagery, via Google AI Blog.

Those who finished season two of The Mandalorian — if you haven’t and intend to, stop reading now — were treated to the epic reveal of Mark Hamill returning as a post-RotJ Luke Skywalker. The moment was somewhat undercut by the digital de-aging effect, which left Luke firmly in the uncanny valley. But before long, a new hope arose. Its name? Shamook. The YouTuber single-handedly improved on the work of an entire FX studio with an expert deepfake, and his video of the improved scene has since logged nearly three million views. Perhaps, then, we shouldn’t be surprised to learn that ILM has turned around and hired him.

Lucasfilm Hired the YouTuber Who Used Deepfakes to Tweak Luke Skywalker ‘Mandalorian’ VFX, via IndieWire.

Project December is a chatbot-style language generation interface that mimics human speech so well it can be disconcerting. It was one of the first ways that regular people could interface with GPT-3, the large language model trained on half a trillion words scraped from the internet. It can’t tell time or add numbers, it’s known for redundancies and incoherence, and it inherently demands regulation. But its capacity for personalization makes it as easy to “chat” with Spock as with a long-missed loved one. SF Chronicle goes long on technologically facilitated grief.

The Jessica Simulation: Love and loss in the age of A.I., via San Francisco Chronicle.

____

YOLOX: Exceeding YOLO Series in 2021

The popular model architecture known as YOLO has officially started using iPhone naming schemes, jumping from YOLOv5 to YOLOX (though this architecture is closer to YOLOv3 than YOLOv5). This newest version of YOLO incorporates modern features such as anchor-free detection and a decoupled model head. The authors make an interesting claim that the combination of MixUp and Mosaic (data augmentation techniques) means they don’t need to pre-train on ImageNet to get their results. The large version of this new YOLO variant won the 1st Place on Streaming Perception Challenge (Workshop on Autonomous Driving at CVPR 2021).

____

Think The Juice was worth the squeeze? Sign up here to receive it weekly.