mrcal 2.3 release notes

Calibrations are still implemented primarily off chessboard observations, but support for discrete points is being added bit by bit:

• Added mrcal.residuals_point(), an analogue to mrcal.residuals_chessboard()
• The mrcal.show_residuals_...() functions now display residuals from chessboard and point observations
• mrcal.show_geometry() and mrcal-show-geometry and mrcal-show-... --observations can display calibration-time points
• Projection uncertainty can be computed with calibrations from point observations if the points were fixed

CAHVORE models now have gradients implemented, so all the normal functionality now works with these models. We can solve for a CAHVORE model.

This model is noncentral, so unprojection only works if \(E = 0\): when this model behaves centrally

• mrcal_image_..._crop() is now static inline. This fixes linker errors on some compilers
• mrcal_cameramodel_t now has double intrinsics[0] instead of double intrinsics[]. This allows #include <mrcal.h> to work inside C++ code
• mrcal_cameramodel_t has a size unknown at compile time, so it cannot be allocated on the stack. To resolve this we now have mrcal_cameramodel_LENSMODEL_XXXX_t structures for models that have known-at-compile-time size. These can be allocated on the stack, so they are easier to use
• Dense stereo processing in C is now possible. stereo.h contains:
  • mrcal_rectified_resolution()
  • mrcal_rectified_system()
  • mrcal_rectification_maps()

• mrcal-stereo writes out a binary disparity image as a 16-bpp .png file. This allows lossless post-processing of the stereo result
• mrcal-stereo --viz stereo displays both the rectified and input coordinates under the mouse cursor
• mrcal-stereo can process multiple images in parallel
• bug fix: mrcal-stereo respects --axis-scale
• bug fix: mrcal-stereo can once again process multiple files with a single invocation

• mrcal.show_projection_diff() and mrcal-show-projection-diff handle nan values better: no confused plots or ugly warnings on the console
• mrcal.show_projection_diff() and mrcal-show-projection-diff can plot vectorfields with multiple distances
• mrcal.show_projection_uncertainty_vs_distance() now has distance_min, distance_max, observed_pixel_uncertainty arguments to better control the plot. The defaults autodetect the values, so the default behavior does not change
• mrcal.show_residuals_board_observation() and mrcal-show-residuals-board-observation reports iobservation_from_worst in its title. This makes it simple to regenerate specific plots, even if they were displayed out of order
• mrcal-show-residuals-board-observation has --from-glob

• mrcal.save_image() can write 16-bit-per-pixel grayscale images properly. This requires a more recent libfreeimage than is available on some OSs. The tests make sure it works.

mrcal.estimate_monocular_calobject_poses_Rt_tocam() ignores invalid input points. This makes calibrating with incomplete chessboard observations work properly

• Commandline tools now print errors to standard error instead of throwing exceptions
• mrcal.rectified_resolution() has been split from mrcal.rectified_system()
• mrcal.optimize() and mrcal.optimizer_callback(): all arguments are keyword-only, most arguments are optional
• mrcal.measurement_index_...() all return <0 if those particular measurements aren't present
• Added mrcal.make_perfect_observations() to produce perfect observations with perfect noise. Very useful for error analysis
• mrcal.ref_calibration_object() can return the position of any point on the chessboard, not just the chessboard corners
• mrcal-convert-lensmodel has --cbmax
• mrcal.residuals_chessboard() argument rename: i_cam \(\rightarrow\) icam_intrinsics for consistency. The old argument is still accepted for backwards compatibility
• mrcal.residuals_chessboard()() has a new return_observations argument The default return_observations=False produces the original behavior. if return_observations: we return residuals,observations
• mrcal.residuals_chessboard() returns size-0 arrays if no observations are present
• mrcal.show_geometry() has smarter logic when given > 2 cameras: the axis_scale has a reasonable default size