Working with the Domain Experts

Learning from Botanists I was recently invited to give a keynote at a biodiversity symposium. In my talk, I covered a brief history of machine-vision computing tools, from SIFT (scaled invariant feature transform) to DPM (deformable parts model) and to DNN (deep neural networks). I talked about the ups and downs of neural networks and their application in digits recognition, face detection, and image classification. I specifically talked about a recent collaboration with botanists in building a large-scale plant recognition system. We reported promising results, and audience members—mostly researchers in biodiversity—were excited to see the significant progress in plant classification and recognition. They were particularly impressed by the power of DNN. What impressed me most during this research collaboration with the botanists was learning about the vertical domain knowledge in plants. Seemingly simple, recognizing plants and flowers can be very challenging. For example, the appearance of a single flower can be different from that of a group of flowers, a top view can be different from a side view, and a flower with leaves and stems can be different from just the flower itself. The botanists taught us that using all parts of a flower can significantly improve classification accuracy.