This uses the same time source as CSCore does for image captures, and will make latency measurements more accurate.
Co-authored-by: Banks T <btrout.dhrs@gmail.com>
Modified VisionModule.java to cause any camera the either has a vendor-defined FOV or the PiCam quirks to control the LED state.
This is a bit of a patch, and directly assumes that piCam's are the devices primarily used for vision processing.
This fixes a tiny bit of pain experienced with a pi3b + picam v1 + usb Cam where not having the USB cam causes the LED's to be working normally, but plugging in the USB cam causes them to be always off.
Parallel issue entered in documentation to add this to known limitations of photon - if you had a set of all usb cameras, we wouldn't know which one is being used for vision processing, and would not be able to control the LED's accordingly. The ultimate solution is just to have teams control the NT entry themselves, which is easy enough. But, docs shouild be updated to reflect that.
* Incorporated new images from gloworm beta unit from Declan into unit test suite.
* Missed one unit test case - looks like the 1280x720 image set also returns wonky center values.
* WIP monkeying around with adding new targets.
* Added testcase to replicate large extrinsics result from calibration.
* Tweaked calibration pipline to return image with chessboard corner detection overlaid.
* Removed "bad" images from the cal, but that didn't seem to help...
* Added test logic to output the undistorted Mat for evaluating the quality of the calibration.
* Tweaked generation of chessboard points to be in a square pattern, not parallelogram.
* Spotless and removed bespoke test target in prep for PR.
* Revised to a double-for loop for less complex-looking logic
* #thanks spotless
al;ksgfjh akljghf ;lakdfdhg ksadfgh klasdfjhg kasdfjghj aklsjg two spaces
* Boop
* Updated findBoardCorners to operate on the image at a fixed, small resolution.
This means the image is scaled down before passing to openCV for processing, then the returned point cloud is scaled back up by that same factor.
Added additional flags to findBoardCorners to "fail fast".
Revised subpixel optimization parameters to use a dynamically-sized window (based on the distance observed between board corners).
* Added additional unit testing on calibration at different resolutions
* Spotless cleanup
* Add Mat release calls
* One more missing release.
Also, tried to make spotless a bit happier
* Additional formatting and WIP tracking down t he memory leak
* Cleaned up `Point` allocation in utility functions that iterate over mats of points
* Maybe fixing this bugger????
* Indeed! We can now get through the image capture stage of calibration at full resolution and save off images. Still crashing on out of memory as expected.
Added additional unit sim pass/fail criteria to catch unreleased mat's and a new set of images for testing high-res cal.
* Letting spotless do its thing
* Undistort debug seems to be acting a bit wonky in CI, so just commenting out for now
* Guard against testcases bleeding state into each other.
In particular, at least in CI, it appears not all mats are getting released in previous testcases, which borks up a hard check against "0 mats allocated" in our calibration tests.
* Removed obsolete tests
* One of these days, I will indeed learn to run spotless before I push.
But today is (still) not that day.
Today we push. FOR ROHAN!
* Cherry-pick the extra debug info from Bank's patches.
* Updated Calibrate3d pipeline to release all unnedded mat's prior to returning.
Update a few raw mat operations to use CVMat for better traceability.
* spotless cleanup
* Added check to shouldPrint for optimization on cvmat deallocate
* Reworked stack trace printing to use lambdas for efficiency
* Missed an unneeded logger.trace
* Formatting improvements
* WIP monkeying around with adding new targets.
* Added testcase to replicate large extrinsics result from calibration.
* Tweaked calibration pipline to return image with chessboard corner detection overlaid.
* Removed "bad" images from the cal, but that didn't seem to help...
* Added test logic to output the undistorted Mat for evaluating the quality of the calibration.
* Tweaked generation of chessboard points to be in a square pattern, not parallelogram.
* Spotless and removed bespoke test target in prep for PR.
* Revised to a double-for loop for less complex-looking logic
* #thanks spotless
al;ksgfjh akljghf ;lakdfdhg ksadfgh klasdfjhg kasdfjghj aklsjg two spaces
* Boop
* Reverting my changes in Calibrate3dPipeline.java to make this merge better with other PR's
* Derp changed the wrong one
Removes duplicate members, try-catches streaming processing, and prevents drawing empty contours. The latter likely stems from some sort of use-after-free condition.
