CNC computer vision updates + pairing ideas

Hi friends,

I went deeper into computer vision this past month, topping last month’s selfie-in-CAD with an 16-bit endian-reversed depth selfie:

Selfie of Kevin from a depth camera

For my CNC computer vision research project, I’m experimenting with posting weekly video updates. So far I’ve done three:

The CNC computer vision problem and manual perspective transform workflow
Automatic machine bed detection using fiducial markers
Generating high resolution image mosaics

The most recent video manages to shave many yaks, including:

Using Rust for low-latency video streaming from a webcam on Windows, over UDP, to a Rust program running within Python on Mac
Controlling a CNC machine via the Windows Registry (!!?!)
Using a tiny bit of JavaScript (but mostly CSS variables) to whip together a GPU-accelerated image stitching analysis tool

Anyway, check out the research log and let me know what you think.

Pairing

What kicked whole computer vision stuff into gear for me was a few video call pairing sessions with a friend. This has reminded me how much I love pairing and that I’d love to do it more, so I’m now keeping a running list of ideas as an open invitation — so if you’re interested in jamming (in person in Oakland or via video call), definitely let me know!

Misc notes

Sounds of the Nightmare Machine: A guitar builder designs a spooky noise machine for a horror movie composer.
Digital projectors rely on tiny mirrors that can operate at 20kHz. But videos are running at, what, 60FPS? So why do the mirrors need to be so much faster? Check out the answer, which includes sweet slow-motion footage of the mirrors in action. (Via The Prepared newsleter.)
This Performance Matters StrangeLoop talk is spectacular. While they don’t describe their method in this language, they essentially devise an LLVM plugin that measures the performance gradient of your program in the space of all its functions — so you can predict that, e.g., that speeding up function X by 10% will improve whole program performance by 17%. They actually use their method to speed up SQLite by 25% (!!!), which is wild, since previously my mental model of SQLite was as a polished, unimprovable perfect sphere of pure software.
Excellent overview of computational photography; gets into light fields, high dynamic range, color correction, and all the techniques by which optically poor smart phone cameras can yield amazing photographs via clever computering.