Machine vision people: given a video of a pcb in a reflow oven, how hard would it be to detect the timestamp at which solder melts on each visible terminal and map it back to the cad data?
@azonenberg general thoughts:
filming with the IR cut filter removed may be a good option here, unless it's an IR heater.
locating the pads is a separate problem to spotting the solder change, and likely the hardest problem here. you may need to do some cleverness to get a polygon mask for each one since I imagine you won't get perfectly cardinal bounds.
once you've got region masks for each pad it's likely a simple image difference scoring problem.
@azonenberg the video compression introduces some hazards but you can largely null those out by doing rolling frame averages. if you only need a timestamp to within a second or two, you can get pretty liberal with the temporal averaging.
@azonenberg solving the pin locations is a bit above my pay grade. the best idea I have is something like manually selecting 4 corners of the board, solving for the distance and angle you'd need to do a KiCAD render for the same perspective as the camera, map the rendered board corners onto the video, then use a multivariate minimiser algorithm like BFGS to tune shear/skew/pinch/punch until you get the closest possible result. then reverse projection map back to the pin locations in the render.
@gsuberland @azonenberg if you already have the four corners, why not just do a planar perspective transform right back onto the layout?
@floe @azonenberg because lenses aren't perfect. you'll get a difference in FoV and some warping. for fine pitch details across the board the inverse projection mapping will be prone to small angle error, so you need to get it as close as possible.
@azonenberg I don't have an answer for you, but I do have the opposite: every modern resource online is going to tell you to use VQA for this. do not use VQA for this, it will be *very* inefficient and unreliable.