Vision Compositing - Multi-Shot Bottom Vision
ma...@makr.zone
Jim Freeman
That is a very interesting video. I have a couple of questions. With the multiple shots, are you building a larger image for the pipeline to use? For example if I do multiple shots, then do the findcountours pipeline stage, will it process the expanded (number of pixels) image file to find the contour that surrounds the full part? What about backlash issues, where you move the part corner to the center of the camera, but in fact the machine does not do that perfectly due to backlash ?
Best, Jim
Sent from Mail for Windows
--
You received this message because you are subscribed to the Google Groups "OpenPnP" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openpnp+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/openpnp/cf5d2aca-3f02-4fb0-8530-8a6a0b3437d9n%40googlegroups.com.
mark maker
> With the multiple shots, are you building a larger image
for the pipeline to use?
No, each shot has its independent detection of a corner or
bracket configuration. All that is consolidated afterwards are
these corner coordinates.
Stitching an image would be very problematic and also only create
a very fragmented image, that is not usable in a vision stage.
> will it process the expanded (number of pixels) image file to find the contour that surrounds the full part?
No, as follows from above.
> What about backlash issues, where you move the part corner to the center of the camera, but in fact the machine does not do that perfectly due to backlash ?
Use backlash compensation:
https://github.com/openpnp/openpnp/wiki/Backlash-Compensation
And if there are still error left, they are likely quite random,
so using multiple corner shots will actually improve accuracy,
simply through the statistic nature of collecting multiple
measurements.
_Mark
To view this discussion on the web visit https://groups.google.com/d/msgid/openpnp/83C1BF34-2A7E-40AC-8751-AC6799228D0E%40hxcore.ol.
Jim Freeman
Thanks, Mark. In the (I guess new or modified) MinAreaRect, is it clear how the new corner points are passed into it?
Best, JIm
To view this discussion on the web visit https://groups.google.com/d/msgid/openpnp/c21ade37-7712-6efc-7a49-b45792fcd229%40makr.zone.
mark maker
Yes, that was a preparatory Pull Request, see here:
To view this discussion on the web visit https://groups.google.com/d/msgid/openpnp/CD51AC2A-3423-4631-B6E3-E42740F76C3E%40hxcore.ol.
Jason von Nieda
vespaman
mark maker
Responding to Kim Lindberg:
> Hi Mark,
Something like this?
https://awesomeopensource.com/project/adamjvr/ESP32-kiCAD-Footprints
You could probably start by generating the left/right pins automatically, as a "Dual" style footprint. Then add the lower pads manually. For multi-shot vision as it is today, it would be sufficient to add one big fused "pad" as a bounding rectangle of the lower pads.
WARNING: the generator creates the side pads symmetrically, and it has to be like that, the multi-shot must find some symmetrical corners. So for that example ESP32 footprint, the center of the must be aligned with the 6 and 26 pads in your E-CAD, i.e. the footprint center in the example would have to be shifted (as I indicated with red arrow and yellow cross).
It is possible that multi-shot vision would one day in the future support Vision Offsets, but not yet. The way this is currently implemented is not sufficient for multi-shot: it currently shifts the center after having it detected asymmetric, it would have to do it before, to make detection symmetric.
