Aruco module, CORNER_REFINE_APRILTAG, speed up?

asked 2018-04-02 12:36:55 -0500

antithing gravatar image

updated 2018-04-02 12:42:33 -0500

I am tracking markers with cv::aruco. I have just updated to the newest version of this module, and i see that there is a new corner refine method, CORNER_REFINE_APRILTAG

Switching to this improves my marker detection a LOT. the range of marker detection has doubled. BUT, at the cost of speed. The detection is around 8 x slower it seems.

My question is, can i tweak the parameters to find a happy medium between robust detection and speed?

I see these options in aruco.hpp under detector params:

- aprilTagMinClusterPixels: reject quads containing too few pixels.
 * - aprilTagMaxNmaxima: how many corner candidates to consider when segmenting a group of pixels into a quad.
 * - aprilTagCriticalRad: Reject quads where pairs of edges have angles that are close to straight or close to
 *   180 degrees. Zero means that no quads are rejected. (In radians).
 * - aprilTagMaxLineFitMse:  When fitting lines to the contours, what is the maximum mean squared error
 *   allowed?  This is useful in rejecting contours that are far from being quad shaped; rejecting
 *   these quads "early" saves expensive decoding processing.
 * - aprilTagMinWhiteBlackDiff: When we build our model of black & white pixels, we add an extra check that
 *   the white model must be (overall) brighter than the black model.  How much brighter? (in pixel values, [0,255]).
 * - aprilTagDeglitch:  should the thresholded image be deglitched? Only useful for very noisy images
 * - aprilTagQuadDecimate: Detection of quads can be done on a lower-resolution image, improving speed at a
 *   cost of pose accuracy and a slight decrease in detection rate. Decoding the binary payload is still
 *   done at full resolution.
 * - aprilTagQuadSigma: What Gaussian blur should be applied to the segmented image (used for quad detection?)
 *   Parameter is the standard deviation in pixels.  Very noisy images benefit from non-zero values (e.g. 0.8).

I have tried changing aprilTagQuadDecimate to 1.2, or 1.5, and this helps, but adds quite a bit of jitter to the pose. Which of these is best to adjust to improve speed, while keeping the maximum accuracy in pose estimation?


edit retag flag offensive close merge delete