imago
Polli’s spiritual predecessor. Rudimentary computer vision pipeline to infer pollination activity from images & video.
Quickly developed to accelerate my Summer 2021 fieldwork, allowing me to collect ~100x more data than I could have manually. The study would have been impossible to conduct with manual sampling with a single part-time field assistant, so I built something to offload the inefficient parts of fieldwork.
Pipeline sensitivity was much more important than computational efficiency, and I didn’t have the time to develop a research-grade detection model. I had the insight that a sufficiently generalized classifier can function as a detector when passed over many overlapping slices of an image. imago recursively applies a general taxonomic classifier to image slices, aggregates & then filters results with per-class NMS, and then infers floral visitations with the reconstructed bounding boxes.
Produced as part of my (unsubmitted) MSc thesis at Kansas State University.
I may eventually open-source this codebase as a historical curiosity, but it is very ugly.