Open lionfish0 opened 1 month ago
I can't test cross-validation -- as for that to work we need 3 photos of the same code location (the idea is we leave out one, and see if we can predict its image location). With just two cameras, leaving out one means that there's often ambiguity about its 3d location - which makes x-validation impossible.
So we need to collect more photos (with the exposure adjusted).
The colour cameras are set with a broad enough aperture they can see the matrix "QR" codes, but the greyscale ones are set too dark...
Greyscale [left] vs Colour [right] photos... (low res for memory processing reasons, as just kept for plotting)
Greyscale and colour tag detections:
There's an API endpoint that lets us set the exposure (see, e.g.
/config/exposure/10
) but we need to set it for calibration, then set it back to the correct value (whatever that is).