Intel Utilizes Computer Vision to Help Penguins Avoid Extinction

2020-01-21

According to a study by the British Antarctic Survey in 2019, the population of Antarctic emperor penguins is facing serious reproductive issues related to climate change, and there is a risk of extinction by 2100. To find a solution, a high-tech company group led by Intel has developed a computer vision solution to help ecologists identify the remaining penguins faster and more accurately than before. The alliance consists of Intel's AI Builders, a carefully curated software vendor ecosystem, Microsoft's AI for Earth initiative, and data science consulting company Gramener. Gramener obtained a corpus of photos of Antarctic penguin colonies from the Penguin Watch program at the University of Oxford, which has collected millions of time-lapse images from camera traps in over 40 locations over the past decade and recruited online volunteers to annotate them. The data was then fed into a convolutional neural network, which retained spatial information while estimating the overall quantity. The AI model workflow consists of two stages. The first stage classifies the images into broad image categories, and the second stage generates maps by estimating density based on inputs from the early stages. This density-based approach successfully addresses most imaging challenges by approximating the number of penguins in different-sized clusters. Intel Xeon Scalable processors were used to train the models in a virtual environment on Microsoft's Azure cloud platform, and then benchmarked using Intel's PyTorch optimization tools. Through this benchmarking and other open-source tools and libraries, Gramener claims to have reduced training time from several days to a few hours. There are multiple challenges to overcome. Soumya Ranjan Mohanty, Chief Data Scientist, said that almost 70% of the images are unusable due to poor lighting or foggy conditions, making the penguins invisible. "These are easily noticeable to the human eye, but difficult for algorithms," he added. This is not Gamener's first AI project with an ecological focus. The consulting company collaborated with Microsoft to build a model using open-source data from iNaturalist, which can classify thousands of different species of animals and plants. They also developed a two-step solution to remove unusable images from camera traps and identify images containing animals, which became Microsoft's camera trap API, as well as a system that provides high-resolution land cover information. Additionally, Gamener worked with the Nisqually River Foundation to select fish by analyzing video footage from underwater cameras that are triggered when fish pass certain infrared sensors. In Kenya, they collaborated with local researchers to develop a model that can distinguish elephants from other livestock in aerial photographs. Of course, Gamener is not the first company to apply AI to such problems. In December 2019, Google collaborated with Conservation International and other organizations to process the world's largest and most comprehensive database of photos taken by motion-activated cameras. DeepMind conducted detailed ecological research with its scientific team last year to develop an AI system that helps study animal behavior in Tanzania's Serengeti National Park. Microsoft recently highlighted Conservation Metrics, a startup based in Santa Cruz, California, that uses machine learning to track African elephants. This is similar to an independent effort by a team of researchers to develop a system that can identify, describe, and count wildlife in Serengeti snapshots with an accuracy of 96.6%. Intel's TrailGuard AI system also prevents poaching by detecting camera movement using AI algorithms.

RECENT AI NEWS

RECENT AI TOOLS