After taking some real duck shots the next step was to recognize ducks on them.
I’ve found that tensorflow is the sexiest thing at the moment. Maybe there are better ways to achieve what I need, but I’m no computer vision expert.
While I was struggling with installing the whole framework and drivers a friend of mine compassionately suggested that maybe it will more efficient to make infrared shots of ducks. There is pi cameras without IR filter. Augmented with visible light filter that would be an easy solution to my problem! I was very encouraged by googling “duck infrared”:
(photo taken from http://coolcosmos.ipac.caltech.edu/image_galleries/ir_zoo/duck.html)
Then we’ve discovered that actually quite wide part of the spectrum is called “infrared” and no way pi camera is heat radiation sensitive. So no easy wins here unfortunately.
Back to tensorflow, that’s what “Faster RCNN with Inception Resnet v2” model thinks of my photos:
While definitely cool, that’s not perfect (not all ducks are recognized), and somewhat slow, as I’ve used the most sophisticated model. Maybe I need to train my own model to recognize ducks specifically.