Benjamin Kellenberger

Species distribution map of the Bighorn Sheep (Ovis canadensis).

You can hover over the image to the right to see one basic example product of what machine learning models can deliver.

Wait, what's so hard about this?

Predicting bounding boxes (i.e., object detection) has long been studied in computer vision, sure. However, aerial imagery poses a completely different set of problems. If you believe this to be simple, I would like to encourage you to get a stopwatch, open up this aerial image and measure the time it takes you to find the mammals. Then, think about how hard it is—even for a computer—to do that in hundreds of thousands of such images with a high recall and low error rate...

Why even bother?

Locating animals in aerial imagery serves many purposes, such as:

• Censuses: wildlife parks, game reserves, and authorities need to periodically obtain animal counts to monitor the health of their livestock.
• Poaching prevention: even to this day, thousands of elephants, rhino, and more are poached every year in Africa. If this continues, much of our great fauna will become extinct. For example, the Black Rhino is about to die out in a few years from now...
• Knowing where the animals are and how many there are helps identifying grazing hotspots and sending out anti-poaching patrols. Plus, on a more civilised note, ecologists have great interest in these data to establish links between various ecosystem parameters, and to study the interactions between animals and flora.
• Doing all of this using aerial images enables covering larger areas from a better point of view (less occlusions). Doing it automatically means that humans need less time to annotate, the localisation quality will improve, and the costs go down.

How does this work?

As is widespread these days, much of my research uses computer vision and machine learning (including, but not limited to, deep learning) to identify and localise animals. See the publications page if you want to know more.
However, simply developing models and publishing them in papers will not automatically change things for the better. Instead, my work is fuelled by two critical stances:

• Machines work best if they assist and collaborate with humans. I oppose full automation of jobs, unless the task is too dangerous or requires immense precision. Computers are good at large-scale, monotonic tasks, so let's offload this portion of the work to them, and leave the final reasoning to humans.
• I strongly believe that scientific research needs to be applied and given to the right audience, academic or not, in order to be successful. I wish to see a direct impact of my work, beyond citations.

You may want to check out my latest software projects and publications.