I wanted to get one more update in before the end of February, luckily we have a leap day this year!
I’m planning a live coding session for That Thing in Swift on Wednesday this week. I still have a few technical things to work out but I’m not committed to it being perfect, just good enough to get a sense of how much interest there is. I think it could be pretty cool.
I need to do some more thinking about how to test theories related to Boundary Layer in the very early stage. I did a bit of research into creating wind tunnels to do proper experimentation and - surprise - there are super shitty versions made for high school science fairs and then models created for final Masters theses. Insert scathing comment on the lack of curiosity in the modern human. I’m still too early to build anything but it’s good to know I can at least create a mid-range experimental rig that essentially doesn’t exist yet.
I neglected to mention another new project last time, Deep Birding , which is an attempt to play with machine learning by classifying the many birds that come to eat seed in the backyard. I think it’ll be fairly straightforward to identify bird type given enough training data but I’d really like to be able to identify individuals. I’m not expecting magic, I could identify individual birds myself given enough time and footage but it would be pretty impressive to be able to do it with machine learning.
Two things stand out as difficult: (1) lots of the example image classifiers available already have the images cropped to the same small size for testing. We’ll need to preprocess frames, looking for areas that have birds and then cropping to our processing size. (2) If we want enough detail to identify individuals, that means a relatively large bird image size. Convolutional processing is super fast with small images and might be prohibitively slow with images large enough to contain enough detail.
I’ve already collected 20 minutes of 1080p video with a few different bird species pecking around. I’m still working through the TensorFlow examples and figuring out how everything works so results are still a ways off.