野性的呼唤 (书): GPUs 如何帮助追踪受威胁的动物物种
这篇文章来自 nvidia.com。原始 url 是: https://blogs.nvidia.com/blog/2018/03/13/wildbook-threatened-species/
以下内容由机器翻译生成。如果您觉得可读性不好, 请阅读原文或 点击这里.
It’s not true that if you’ve seen one zebra you’ve seen them all. But unless researchers can get a handle on their real numbers, it one day may be.
To do that, you have to be able to identify individual zebras and — as Tanya Berger-Wolf discovered after spending 25 minutes trying to identify a specific zebra — that’s even harder than it sounds.
A zebra’s stripes are as subtly individual as human fingerprints. “I thought I was going to lose my mind,” admits Berger-Wolf, a professor of computer science at the University of Illinois at Chicago.
So Berger-Wolf helped create a solution that harnesses publicly shared videos — and GPUs — to speed up the task of identifying individual animals of all kinds.
Called Wildbook, the tool promises to solve a big headache for Berger-Wolf, who runs her school’s Computational Population Biology Lab, and the many wildlife researchers she works with.
Take the challenge of identifying zebras.
Berger-Wolf found her software’s computer vision component was great at identifying zebras. And she expanded to other species by partnering with Charles Stewart, a computer vision researcher at the Rensselaer Polytechnic Institute. But matching images of individual animals with others within a dataset was a slow, immature process.
During her subsequent search for options, Berger-Wolf came upon the nonprofit WildMe, which had employed a superior data model to create a log of sightings of whale sharks. She figured her team and the WildMe staff could help each other.
“They were interested in adding computer vision to the system, and we were interested in a mature data management layer,” Berger-Wolf said. “It was love at first sight.”
The resulting Wildbook project is a multi-organizational effort to protect and preserve endangered and threatened animals.
Using publicly and intentionally shared images and videos, the Wildbook team creates species-specific visual databases of encounters with animals, incorporating whatever data can be gleaned from the photos and videos.
Extracting Data from Videos
To accomplish this, videos from YouTube and social media sites are scraped using AI-infused computer vision to identify each animal. Natural language processing is also applied to transcribe any text or audio.
To date, the team has created 15 Wildbooks for animals ranging from giraffes 和 polar bears 自 whale sharks 和 manta rays. It has a backlog of more than 200 requests from conservationists for more Wildbooks.
GPUs figure prominently in the computer vision pipeline, helping the Wildbook team sift through images to classify species. The combination of GPUs and deep learning approaches enables Wildbook creators to dig deep into the data inherent in the images.
“For each image, we not only identify the species, we identify the individual,” Berger-Wolf says. “We can tell you this is a zebra pixel and this is background. We can identify occlusion and classify the quality of the image.”
GPUs are speeding the process significantly. Berger-Wolf estimates GPUs have reduced the time it takes to detect species from a few seconds per image to a fraction of a second — an important consideration as she deals with thousands of images.
Berger-Wolf has lofty goals for the Wildbook project. She sees it becoming a global resource that will help conservationists respond to poaching attacks, or provide scientists with more information on how different species interact with each other.
“We’d really like to expand it to a planetary scale,” she said. “I want to scale it to the level of habitats, to the level of continents.”
Perhaps even more important, she wants Wildbooks to help humans understand how their activity impacts animal environments. Until now, this information has been elusive.
Berger-Wolf notes, for instance, that as of a few months ago, scientists estimated that there were 4,000-6,500 snow leopards in the world, which isn’t real useful information since the uncertainty is too high. Elsewhere, conservationists spent $8 million over a two-year period trying to count elephants so that they could begin to understand the true impact of poaching.
Once Wildbooks are established for those animals, the resulting influx of data will give biologists and conservationists a better understanding of the challenges they face in preserving them. Eventually, Berger-Wolf wants them to be able to interact with the data dynamically.
And this is only possibly with GPU acceleration, considering the large amount of data and high image, spatio-temporal and individual animal resolution.
“I want to be able to zoom in on a map and say, who’s there? How did it change with the difference in rainfall, or with the ocean currents?” she said. “We can really enrich our knowledge of basic information about species.”