The tale of the bee-saving christmas tree

The tale of the bee-saving christmas tree

It is the time of the year gifts are exchanged, candles are lit and fireplaces warm our hearts. In our case, Santa got in a little early and brought about 64 NVIDIA Jetson Nano dev kits. We used them to build the Jetson tree with a total of 8.192 CUDA cores and 256 CPU cores. It’s flickering LEDs now light our office, warming both our hearts and our neighborhood. We will use its power to analyze bee videos in order to investigate the reasons behind the perishing of insects.

External Media

At apic.ai, we believe technology can help us create a better understanding of nature. In our case, we use it to find the reasons for the perishing of insects. We do so by analysing bee behavior like motion patterns and pollen intake. Our monitoring system visually detects bees as they enter and leave their hives. Through their level of activity, mortality and food abundance we gain insights into the well-being of the insects and the plant diversity in the environment. The information thus enables us to evaluate the regional living conditions for insects, detects problems and propose measures to improve the situation.

This winter, we are upgrading our hardware platform from the Raspberry Pi to the NVIDIA Jetson Nano. To do so, we needed a way to test our software on a large number of devices. Our Jetson Tree consisting of 64 NVIDIA Jetson Nanos, will be processing multiple terabytes of video footage that we recorded at bee hives in 2019. While being fashionably in season, it enables us to better understand the Jetson Nanos behaviour under load and resolve bugs. Thus we can have a clean start into 2020’s bee season, when they become part of our monitoring systems and have to perform at remote bee hives.
As a bonus we now have our own GPU cluster to run experiments on.

How did we build it?
Even though this seems to be a simple project, it is a time-consuming build. Due to the sheer size of the cluster it took us a week to finish.

We decided to CNC mill a wooden tree and attached 8 Jetsons to each side. To test different software stacks, we used Open Balena to manage the device fleet and deploy software. Since we are currently using Mender.io for our fleet management, this was an excellent chance to compare containerized software deployment to full image updates.

This build guide covers the following aspects:

• How to build a GPU cluster from Jetson Nanos
• How to setup Open-Balena for fleet management and software rollouts
• How to deploy custom code to the cluster

The Build guide, images and a more detailed description can be found here https://github.com/apic-ai/jetson-tree

Looking forward to get some feedback on our build and ideas on what we should do with it :)

Hi!

Very interesting application, I must admit.

Few questions, though.

What is the evaluation model of the problem? How and what will you measure and compare, as soon as you are done recognizing the collected video footage?

Was that confirmed that a bee that will die soon changes its behavior (and therefore you can measure or sense it)? Imagine that there is a predator killing them. The bees will keep behaving normally till they are killed. What will the image recognition bring in this case?

Regards

Hi Opaliy

The honey bee plays a key role in our approach to determining the causes of biodiversity loss. The colonies live in a relationship of mutual dependence with the flora in their environment. Where the nature around them is not diverse, they do not find food all year round and the development of the population is negatively affected. If the environment is polluted with environmental toxins, the number of bees, and generally the number of insects, decreases. Thus, the condition of the environment is reflected in the honey bees.
Other insects are equally harmed by a toxic environment and lack of food, but the effects cannot be observed as easily. The symbiotic relationship of the honeybees with nature, coupled with the good conditions to observe them in their hives, makes them a unique natural sensor for us.

Image recognition enables us here to count in and out going bees and therefore determine the loss of honey bees and extract further single entity features.
You can find more here: https://apic.ai/

Regarding evaluation models, we are using a custom bee detection network that is currently able to identify and track more than 200 single bee instance simultaneously. Further assessments are done in the cloud with the help of our cool multimodal network DeepBees. presented at ICCV this year: http://openaccess.thecvf.com/content_ICCVW_2019/html/CVWC/Marstaller_DeepBees_-_Building_and_Scaling_Convolutional_Neuronal_Nets_For_Fast_ICCVW_2019_paper.html

Best regards,
Frederic