Kit Quitmeyer
- @wellreadpanda
- | she/her
Hello! Kit Quitmeyer here, co-founder of the Digital Naturalism Laboratories (Dinalab), where our primary project of the past few years has been the Mothbox. The Mothbox is an open source, low-cost, automated light trap that attracts, photographs, and uses computer vision to identify insects, which has received most of our funding from WILDLabs and ARM. The global decline in biodiversity has been recognized as one of the greatest challenges faced by our planet today, so our broader project's mission was to make insect biodiversity monitoring accessible at a landscape-scale (100s to 1000s of hectares).
I was very pleased to receive funding from the Boring Fund in order to improve our computer vision software, which I will refer to as the Mothbot from here. In particular, my proposal aimed to improve the Mothbot's detection model, which isolates and crops images of individual insects from raw image files of an entire Mothbox target with many insects. In this way, the Mothbot saves users from having to manually examine hundreds of images, many of which might not include any insects at all, look for the individual insects, and isolate the insect themselves for identification. Instead, they can run our scripts and process an entire night's data in 15 minutes automatically (which would likely take 5 hours if done manually). As important and useful as this work is, it is VERY boring, requiring me to look through thousands of images and draw rectangles around tens of thousands of insects and other creatures.
A related aim of this funding period was to train the Mothbot on raw images from a broader geographical range. In the past, I mostly trained the detection model on data collected in Panama. Although there is nothing wrong with the Panamanian data, we want users from all over the globe to be able to detect insects worldwide. To this end, we asked global Mothbox users to share their raw images with us. Thanks to their data, during this period I have been training the Mothbot on images taken from countries such as Canada, Indonesia, Hawaii, Poland, China, and Peru, which makes the model more robust.
Journey through the Grant Period
I had proposed spending twenty hours drawing rectangles around creatures and tweaking the rectangles drawn by the detection model. My goal was to process 1,500 images this way. Although I spent more like forty hours on this work during the funding period, I only processed about 500 images so far. I do not consider this a real shortfall, given that the goal of 1,500 images was pretty ambitious. Additionally, it can be hard to estimate how long it may take to process any individual image, as some photos show only a few bugs, whereas others can capture well over a hundred in one picture. Obviously, it takes a lot longer to properly crop images of a hundred bugs than just a few.
Achievements and Outcomes
The Mothbox team is currently on a "world tour" where we are attending conferences and visiting various teams throughout the world who are using the Mothbox and/or the Mothbot. For example, I write to you from Aarhus University in Denmark, where teams are using either the Mothbox for insect monitoring or the Mothbot as supplemental software for processing images captured with other insect monitoring devices. We have been receiving lots of positive feedback about the usefulness of the software in particular, which is a great motivator to improve it even further.
While other members of our team work on the Classifier portion of the Mothbot, I am concentrating on continuing to improve our detection model by processing an additional 1,000 images to meet my initial goal of processing 1,500 images. It is hard to tell how long this might take, but I estimate at least eighty more hours of work.
Once I achieve that number of processed images, I will feed all of this information back into the Mothbot to improve its detection model. Additionally, I will publish an open access database of the ground truth based on a global dataset on the Zenodo platform so that anyone can use that data.
Add the first post in this thread.