Thanks to this tutorial from TheCodingBug, I was able to use YOLOv7 to create a very primitive bird detection network. I scraped some generic Pennsylvania bird images, labelled them with a single “bird” class, and let YOLO work its magic overnight.
For only an evening’s worth of work, I am extremely happy with the results. No–it’s not perfect, but it’s REALLY promising. With the amount of footage I have recorded and can potentially label in time, I have no doubt that I’ll be able to fine tune this model for our feeders and improve its recall even more.
What’s more incredible to me is just how straightforward it was to set something like this up. I’m not a machine learning expert. I know the very basics and fundamental concepts–that’s about it. To be able to train my own object detection network in a night without much of any prior experience is a testiment to how powerful these machine learning frameworks are for putting this technology in the hands of anyone who needs it.
Now, it’ll take quite a bit of learning to really figure out the details of improving this, but I’m just so excited I could get this part of the project started. I’ll try to update this post as I make changes.