Gramener Helps Researchers Recommend Environmental Conservation Measures To Preserve Penguin Populations
Penguin populations in Antarctica are diminishing at an alarming rate. Regardless of the cause – climate change, human encroachments or other factors – researchers are racing to reverse the trend, and they’re leveraging artificial intelligence to derive an accurate count of the remaining birds.
By employing camera traps to capture millions of time-lapse images, environmental scientists have collected rich data about the decline of penguin colonies. However, the challenge has been in counting the millions of birds captured in images spanning the last 10 years. Moreover, those penguin colony images suffer from the same crowd-counting issues as pictures of massive human gatherings, namely:
- Perspective distortion, where faces in the front appear larger than the ones in the back.
- Occlusion, or individuals in the back being hidden from view.
- Density differences within the same picture, where some portions have unusually high clusters of individuals.
- Diversity of camera angles, where top-down pictures need separate training than frontal angles. Additionally, the penguin colony images also suffer from poor image quality due to adverse weather conditions.
Now comes the solution: a deep learning application for detecting and counting penguins co-developed by data science company Gramener, an independent software vendor (ISV) and a member of the Intel® AI Builders program. In partnership with Microsoft’s AI for Earth initiative, Gramener’s approach – based on convolutional neural networks (CNNs) using density-based estimations – has delivered a more reliable and accurate approach to penguin counting. CNN algorithms preserve spatial information and can localize the count while also estimating the overall tally. Cascaded CNNs have a two-pronged approach, wherein the first stage classifies the image into a broad category and the second stage estimates density by taking inputs from earlier stages to generate more refined maps as shown in the diagram below. Learn more about CNNs for crowd counting.
Image source: https://www.infoq.com/news/2019/04/kesari-crowd-counting-ai/
The theme of doing ‘AI for good’ is of special interest for Gramener. Apart from applying AI for business benefits, they have used it for biodiversity conservation and worked in areas such as monitoring salmon, detecting wild elephants, and identification of biodiversity in other AI for Earth initiatives. Gramener’s work will help enable researchers to recommend ecological and environmental conservation measures to preserve penguin populations.
The Gramener team was inspired by Penguin Watch, an initiative at Oxford University, which collected the data and then used crowdsourcing by citizen scientists to label it. (See this paper from Oxford University researchers for reference.) Volunteers from around the world placed markers on thousands of pictures, which helped the deep learning models learn what penguins look like and how to differentiate them from other image elements. Given Gramener’s longstanding interest in conservation initiatives, they took up this challenge of penguin identification and estimation. View a penguin-counting demo.
In approaching this problem, Gramener had several training methods to choose from. For example, traditional deep learning methods scan the image to identify, say, heads. It estimates the total count by drawing bounding boxes around all such matches. However, this method presents certain disadvantages: for example, penguins in the back, hidden from view by other penguins cannot be identified. Density-based counting approaches can handle most of these challenges by approximating the number of penguins in clusters of different sizes.
The Intel® AI Technology Behind the Achievement
Intel® Xeon® Scalable processors and optimized frameworks were used for the daunting assignment which utilized a huge dataset. The solution was originally implemented on a virtual machine powered by Microsoft Azure*. It was then benchmarked using Intel® AI technology and the Intel® Optimization for PyTorch*. This configuration helped Gramener train the algorithm at scale, achieving the desired performance at a potentially lower computational cost.
From its benchmarking exercise, Gramener was able to reduce training time from days to hours using Intel Optimization for Pytorch, OpenCV, and other tools and libraries optimized for Intel® Xeon® Scalable processors. Gramener intends to use the same Intel-optimized frameworks and hardware for future projects and will share it as a recommendation to other clients.
Additional Use Cases
Anonymous crowd counting can be applied in a variety of scenarios – not just with penguins. Use cases include:
- Counting People – The retail industry could use this application to potentially get reliable counts of store footfalls, estimate conversions and measure success of campaigns.
- Counting Biological Cells – Drug characterization is a key step in the pharma drug discovery process as scientists need counts of different cell types from microscopic images. This painful, manual process could take advantage of this application to automate the process.
Gramener has recently made its Gramex data science platform available as open source, with components for data handlers, computational and analytics modules and a rich library of interactive visualization charts. The combined power of Gramex components allows data scientists to build highly interactive data applications by configuring required components programmatically. Learn more on Gramener’s GitHub*.
Founded in 2010, Gramener is a data science consulting company that advises clients in data-driven leadership – specifically in data analytics, design and data visualization. Gramener’s services help business users acquire insights through analysis of data, leveraging machine learning, artificial intelligence and automated analysis, and through visual intelligence via modern charts and narratives.