mrcal-calibrate-cameras - Calibrate some synchronized, stationary cameras


  $ mrcal-calibrate-cameras
      --corners-cache corners.vnl
      --lensmodel LENSMODEL_OPENCV8
      --focal 1700 --object-spacing 0.01 --object-width-n 10
      --outdir /tmp
      'left*.png' 'right*.png'

    ... lots of output as the solve runs ...
    RMS reprojection error: 1.9 pixels
    Worst reprojection error: 7.8 pixels
    Noutliers: 319 out of 17100 total points: 1.9% of the data

    Wrote /tmp/camera0-0.cameramodel
    Wrote /tmp/camera0-1.cameramodel


This tool uses the generic mrcal platform to solve a common specific problem of N-camera calibration using observations of a chessboard. Please see the mrcal documentation at for details.



  images                A glob-per-camera for the images. Include a glob for
                        each camera. It is assumed that the image filenames in
                        each glob are of of the form xxxNNNyyy where xxx and
                        yyy are common to all images in the set, and NNN
                        varies. This NNN is a frame number, and identical
                        frame numbers across different globs signify a time-
                        synchronized observation. I.e. you can pass
                        'left*.jpg' and 'right*.jpg' to find images
                        'left0.jpg', 'left1.jpg', ..., 'right0.jpg',
                        'right1.jpg', ...


  -h, --help            show this help message and exit
  --lensmodel LENSMODEL
                        Which lens model we're using. This is a string
                        "LENSMODEL_....". Required unless we have a --seed.
                        See L<>
                        for notes about how to select a model
  --focal FOCAL         Initial estimate of the focal length, in pixels.
                        Required unless --seed is given. See
                        L<> for
                        notes about how to estimate a focal length
                        Size of the imager. This is only required if we pass
                        --corners-cache AND if none of the image files on disk
                        actually exist and if we don't have a --seed. If we do
                        have a --seed, the --imagersize values must match the
                        --seed exactly
  --outdir OUTDIR       Directory for the output camera models
  --object-spacing OBJECT_SPACING
                        Width of each square in the calibration board, in
  --object-width-n OBJECT_WIDTH_N
                        How many points the calibration board has per
                        horizontal side. If omitted we default to 10
  --object-height-n OBJECT_HEIGHT_N
                        How many points the calibration board has per vertical
                        side. If omitted, we assume a square object, setting
  --seed SEED           A comma-separated whitespace-less list of camera model
                        globs to use as a seed for the intrinsics and
                        extrinsics. The number of models must match the number
                        of cameras exactly. Expanded globs are sorted
                        alphanumerically. This is useful to bootstrap the
                        solve or to validate an existing set of models, or to
                        recompute just the extrinsics or just the intrinsics
                        of a solve. If omitted, we estimate a seed. Exclusive
                        with --focal. If given, --imagersize is omitted or it
                        must match EXACTLY with whatever is in the --seed
  --jobs JOBS, -j JOBS  How much parallelization we want. Like GNU make.
                        Affects only the chessboard corner finder. If we are
                        reading a cache file, this does nothing
  --corners-cache CORNERS_CACHE
                        Path to the corner-finder results. If this file
                        exists, I use the corners in this file. If it doesn't
                        exist, I invoke mrgingham to compute the corners, and
                        I write the results to that path. And THEN I compute
                        the calibration off those observations. This file is a
                        vnlog with legend "# filename x y level" (exactly what
                        mrgingham reports). Each rown is an observed corners.
                        If an image had no observations, a single row
                        "filename - - -" is expected. The "level" is the
                        decimation level used in detecting that corner. "0"
                        means "full-resolution", "1" means "half-resolution",
                        "2" means "quarter-resolution" and so on. A level of
                        "-" or <0 means "skip this point". This is how
                        incomplete board observations are specified. A file
                        with a missing "level" column will fill in "0" for all
                        corners. A non-mrgingham grid detector may be used by
                        running that separately, and using this option to read
                        the output. A detector may output weights instead of a
                        decimation level in the last column. Pass --corners-
                        cache-has-weights to interpret the data in that way
                        By default the corners we read in --corners-cache have
                        columns "filename x y level". If the last column is a
                        weight instead of a decimation level, pass this
                        option. This is useful to support non-mrgingham
                        chessboard detectors
  --pairs               By default, we are calibrating a set of N independent
                        cameras. If we actually have a number of stereo pairs,
                        pass this argument. It changes the filename format of
                        the models written to disk (cameraPAIR-
                        INDEXINPAIR.cameramodel), and will report some
                        uncertainties about geometry inside each pair.
                        Consecutive cameras in the given list are paired up,
                        and an even number of cameras is required
                        By default we apply regularization in the solver in
                        the final optimization. This discourages obviously-
                        wrong solutions, but can introduce a bias. With this
                        option, regularization isn't applied
                        By default we throw out outliers. This option turns
                        that off
                        Keep the seeded extrinsics, if given. Allowed only if
                        Keep the seeded intrinsics, if given. Allowed only if
                        By default we assume the calibration target is
                        slightly deformed, and we compute this deformation. If
                        we want to assume that it is flat, pass this option.
                        For convenience we compute a valid-intrinsics region
                        to describe the results of the calibration. This is a
                        watered-down interpretation of the projection
                        uncertainty that is easy to interpret. The logic
                        computing this is somewhat crude, and may go away in
                        the future. The defaults should be reasonable, so if
                        in doubt, leave these alone. The intent is to produce
                        usable output even if we're using a lean lens model
                        where the computed uncertainty is always overly
                        optimistic. We bin the observations into a grid, and
                        use mrcal._report_regional_statistics() to get the
                        residual statistics in each bin. We then contour the
                        bins to produce the valid-intrinsics region. If we're
                        using a rich lens model (LENSMODEL_SPLINED_...), then
                        we only look at the uncertainty, and not at the other
                        statistics. This argument takes 5 parameters. The
                        uncertainty is computed at a range
                        valid_intrinsics_region_parameters[4]. If <= 0, I look
                        out to infinity. The default is 0. A region is valid
                        only if the projection uncertainty <
                        valid_intrinsics_region_parameters[0] *
                        observed_pixel_uncertainty. The default is 1. A region
                        is valid only if the mean-abs-residuals is <
                        valid_intrinsics_region_parameters[1] (only for lean
                        models). The default is 0.5. A region is valid only if
                        the residuals stdev is <
                        valid_intrinsics_region_parameters[2] *
                        observed_pixel_uncertainty (only for lean models). The
                        default is 1.5. A region is valid only if it contains
                        at least valid_intrinsics_region_parameters[3]
                        observations (only for lean models). The default is 3.
  --verbose-solver      By default the final stage of the solver doesn't say
                        much. This option turns on verbosity to get lots of
                        diagnostics. This is generally not very useful to end
  --explore             After the solve open an interactive shell to examine
                        the solution



Dima Kogan, <>


Copyright (c) 2017-2021 California Institute of Technology ("Caltech"). U.S. Government sponsorship acknowledged. All rights reserved.

Licensed under the Apache License, Version 2.0 (the "License"); You may obtain a copy of the License at