Computer vision technology is about to undergo a major transformation

Will computer vision reinvent itself again?

Ryad Benosman, a professor of ophthalmology at the University of Pittsburgh and an adjunct professor at CMU’s Robotics Institute, thinks so. As one of the founders of event-based vision technology, Benosman expects neuromorphic vision—computer vision based on event-based cameras—to be the next direction in computer vision.

"Computer vision has been reinvented many, many times," Benosman said. “I’ve seen it reinvented at least twice.”

Benosman cited the shift from image processing with a bit of photogrammetry to geometry-based methods in the 1990s, and then today’s rapid advances in machine learning . Despite these changes, modern computer vision technology is still primarily based on image sensors—cameras that produce images similar to those seen by the human eye.

According to Benosman, the image sensing paradigm will hinder innovation in alternative technologies until it is no longer useful. The development of high-performance processors (e.g., GPUs) delays the need to find alternative solutions, thus prolonging this impact.

"Why do we use images for computer vision? That's the million-dollar question," he said. "We have no reason to use images - it's just because of historical momentum. Even before there were cameras, images had momentum." Image cameras have been around since the 1500s, with artists using room-sized devices to trace images of people or scenery outside a room onto canvas. Over the years, the paintings were replaced with film to record the images. Innovations such as digital photography eventually made it easy for image cameras to become the basis of modern computer vision technology.

However, Benosman believes that computer vision technology based on image cameras is extremely inefficient. His analogy was the defense system of a medieval castle: guards positioned around the walls looked around for approaching enemies. The drummer beat steadily, and with every beat, each guard shouted out what they saw. How easy is it to overhear a guard spotting an enemy at the edge of a distant forest amid all the commotion?

The 21st century hardware equivalent of a drum beat is an electronic clock signal, and the guards are pixels. A large amount of data is created and must be checked every clock cycle, which results in a large amount of redundant information and thus requires a lot of unnecessary calculations.

"People are burning so much energy, it's taking up Castle's entire computing power to protect themselves," Benosman said. If an interesting event is discovered - represented by the enemy in this analogy - "you have to move around collecting useless information, people are screaming everywhere, so there is a lot of bandwidth... Now imagine you have a complex castle. All These people all have to be heard."

Enter Neuromorphic Vision. The basic idea is inspired by the way biological systems work, which is to detect changes in scene dynamics rather than continuously analyzing the entire scene. In our castle analogy, this means keeping the guards quiet until they see something of interest, then calling out their location to raise the alarm. In electronic form, this means letting individual pixels determine whether they see something relevant.

"Pixels can decide for themselves what message they should send," Benosman said.

"Instead of getting system information, they can look for meaningful information - features. That's what makes the difference."

Prophesee vs. DVS sensor evaluation kit developed in collaboration with Sony. Benosman is the co-founder of Prophesee.

This event-based approach can save significant power and reduce latency compared to fixed-frequency system acquisition.

"You want something more adaptive, and that's what the relative changes [event-based vision] gives you - adaptive acquisition frequency," he said. "When you look at amplitude changes, if something is moving very fast, we're going to get a lot of samples. If something's not changing, you're going to get almost zero, so you're adjusting your acquisition frequency based on the dynamics of the scene . That's what it brings. That's why it's a good design."

Benosman entered the field of neuromorphic vision in 2000, convinced that advanced computer vision would never work because images were not The correct way.

“The biggest shift is to say that we can do vision without grayscale and without images, which was heretical in the late 2000s — totally heretical,” he said.

The technique Benosman proposed—the basis for today’s event-based sensing—was so different that papers submitted to the most important IEEE computer vision journal at the time were rejected without review. In fact, it wasn’t until the development of the Dynamic Vision Sensor (DVS) in 2008 that the technology began to gain momentum.

Neuroscience Inspiration

Neuromorphic technologies are technologies inspired by biological systems, including the ultimate computer: the brain and its neurons, or computational elements. The problem is that no one fully understands how neurons work. While we know that neurons respond to incoming electrical signals called spikes, until recently, researchers described neurons as rather hasty, assuming that only the number of spikes mattered. This hypothesis persisted for decades, but recent work has proven that the timing of these spikes is absolutely critical and that the brain is structured to create delays in these spikes to encode information.

Today’s spiking neural networks simulate the spikes seen in the brain and are simplified versions of the real thing—usually binary representations of the spikes. "I receive a 1, I wake up, I calculate, I sleep," Benosman explained. The reality is much more complex. When a spike arrives, the neuron starts integrating the value of the spike over time; neurons also leak, meaning the results are dynamic. Additionally, there are approximately 50 different types of neurons with 50 different integration profiles.

Current electronic versions lack integrated dynamic paths, connectivity between neurons, and different weights and delays. "The problem is that to make a product that works, you can't [imitate] all the complexity because we don't understand it," he said. "If we had a good theory of the brain, we would solve it. The problem is, we just don't know."

Bensoman runs a unique lab dedicated to understanding the mathematics behind cortical computation, aiming to before creating new mathematical models and replicating them into silicon devices. This involves direct monitoring of spikes from real retinas.

Currently, Bensoman opposes faithfully replicating biological neurons, calling the approach outdated.

"The idea of ​​replicating neurons in silicon came about because people looked at transistors and saw a mechanism that looked like a real neuron, so there was some thought behind it in the beginning," he said. "We don't have cells; we have silicon. You need to adapt your computing substrate, not the other way around... If I know what I'm computing and I have the chip, I can optimize this equation and do it at the lowest cost, lowest power consumption, The lowest latency to run it."

Processing Power

The realization that exact replicas of neurons are not needed, and the development of DVS cameras, are the driving forces behind today's vision systems. While systems are already commercially available, progress is needed before fully human-like vision can be used commercially.

Benosman said the original DVS cameras had "large, thick pixels" because the components surrounding the photodiodes themselves greatly reduced the fill factor. While investments in developing these cameras have accelerated the technology, Benosman made it clear that today's incident cameras are simply improvements on original research equipment developed back in 2000. The most advanced DVS cameras from Sony, Samsung and Omnivision have tiny pixels that incorporate advanced technologies like 3D stacking and reduce noise. Benosman's concern is whether the types of sensors used today can successfully scale.

"The problem is, once you increase the number of pixels, you get a lot of data because you're still very fast," he said. "You could probably still process it in real time, but you'd get too much relative change from too many pixels. That's killing everyone right now because they see the potential, but they don't have the right processors to support it. ”

This Prophesee customer application example shows the difference between the image camera (upper left corner of each box) and DVS sensor output.

General purpose neuromorphic processors lag behind their DVS camera counterparts. Efforts by some of the industry’s biggest players (IBM Truenorth, Intel Loihi) are still ongoing. The right processor and the right sensor will be an unbeatable combination, Benosman said.

“[Today’s DVS] sensors are extremely fast, have ultra-low bandwidth, and have high dynamic range so you can see indoors and outdoors,” Benosman said. "This is the future. Is it going to take off? Absolutely." "Whoever can put the processor in there and deliver the full stack is going to win because it's going to be unbeatable," he added road.

The above is the detailed content of Computer vision technology is about to undergo a major transformation. For more information, please follow other related articles on the PHP Chinese website!