What is mrcal for anyone who hasn't used it? The line from the website is not making it clearer: "mrcal is a general-purpose toolkit that is useful in solving a variety of problems." What tools? What problems? Image problems?
The mentions of stereo and triangulation intrigued me, but some kind of overview or list of applications at the top of the site would really help.
The most common application is camera calibration (making a model of what the lens is doing). Existing tools (opencv and friends) aren't very good. Downstream of that, anything that uses the calibration results to do stuff (usually either tracking the motion of the camera and/or making maps of what the camera is seeing) can make use of this toolkit also.
Currently the actual stuff you would do with the calibration (traking or mapping) is done by other libraries. The focus here is to produce a more precise, accurate calibration.
Awesome, so if I'm understanding correctly it's analogous to, but better than, the lens modeling that a panorama stitcher will use, and its applications are as the calibration stage of an image stitching, 3D vision, object tracking, SLAM, structure from motion, and/or image-based measurement system?
I've been fascinated by these things since high school and as a teenager tried to make my own object tracking system for special effects for a student film (and obviously failed), but apart from running a Kinect-based home/art/museum automation startup for a while, my life path did not lead me into computer vision.
I have yet to try this out, but the interesting thing in these docs is principled and data supported description of “how to move a calibration target” when calibrating a fisheye camera to get the most bang for buck. This could save companies real money in production line time.
5 comments
[ 4.8 ms ] story [ 11.2 ms ] threadThe mentions of stereo and triangulation intrigued me, but some kind of overview or list of applications at the top of the site would really help.
Given pictures and lens distortion map, the tool can:
* calculate camera position from pair of pictures
* do feature triangulation from pair of pictures - get real-world distance to a selected feature
* calculate depth map from pair of pictures
Currently the actual stuff you would do with the calibration (traking or mapping) is done by other libraries. The focus here is to produce a more precise, accurate calibration.
I'll clarify the docs
I've been fascinated by these things since high school and as a teenager tried to make my own object tracking system for special effects for a student film (and obviously failed), but apart from running a Kinect-based home/art/museum automation startup for a while, my life path did not lead me into computer vision.