mrcal-stereo - Stereo processing
$ mrcal-stereo \
--az-fov-deg 90 \
--el-fov-deg 90 \
--sgbm-block-size 5 \
--sgbm-p1 600 \
--sgbm-p2 2400 \
--sgbm-uniqueness-ratio 5 \
--sgbm-disp12-max-diff 1 \
--sgbm-speckle-window-size 200 \
--sgbm-speckle-range 2 \
--outdir /tmp \
left.cameramodel right.cameramodel \
left.jpg right.jpg
Processing left.jpg and right.jpg
Wrote '/tmp/rectified0.cameramodel'
Wrote '/tmp/rectified1.cameramodel'
Wrote '/tmp/left-rectified.png'
Wrote '/tmp/right-rectified.png'
Wrote '/tmp/left-disparity.png'
Wrote '/tmp/left-range.png'
Wrote '/tmp/points-cam0.vnl'Given a pair of calibrated cameras and pairs of images captured by these cameras, this tool runs the whole stereo processing sequence to produce disparity and range images and a point cloud array.
mrcal functions are used to construct the rectified system. Currently only the OpenCV SGBM routine is available to perform stereo matching, but more options will be made available with time.
The commandline arguments to configure the SGBM matcher (--sgbm-...) map to the corresponding OpenCV APIs. Omitting an --sgbm-... argument will result in the defaults being used in the cv2.StereoSGBM_create() call. Usually the cv2.StereoSGBM_create() defaults are terrible, and produce a disparity map that isn't great. The --sgbm-... arguments in the synopsis above are a good start to get usable stereo.
The rectified system is constructed with the axes
- x: from the origin of the first camera to the origin of the second camera (the baseline direction)
- y: completes the system from x,z
- z: the mean "forward" direction of the two input cameras, with the component parallel to the baseline subtracted off
The active window in this system is specified using a few parameters. These refer to
- the "azimuth" (or "az"): the direction along the baseline: rectified x axis
- the "elevation" (or "el"): the direction across the baseline: rectified y axis
The rectified field of view is given by the arguments --az-fov-deg and --el-fov-deg. At this time there's no auto-detection logic, and these must be given. Changing these is a "zoom" operation.
To pan the stereo system, pass --az0-deg and/or --el0-deg. These specify the center of the rectified images, and are optional.
Finally, the resolution of the rectified images is given with --pixels-per-deg. This is optional, and defaults to the resolution of the first input image. If we want to scale the input resolution, pass a value <0. For instance, to generate rectified images at half the first-input-image resolution, pass --pixels-per-deg=-0.5. Note that the Python argparse has a problem with negative numbers, so "--pixels-per-deg -0.5" does not work.
The input images are specified by a pair of globs, so we can process many images with a single call. Each glob is expanded, and the filenames are sorted. The resulting lists of files are assumed to match up.
There are several modes of operation:
- No images given: we compute the rectified system only, writing the models to disk
- No --viz argument given: we compute the rectified system and the disparity, and we write all output as images on disk
- --viz geometry: we compute the rectified system, and display its geometry as a plot. No rectification is computed, and the images aren't used, and don't need to be passed in
- --viz stereo: compute the rectified system and the disparity. We don't write anything to disk initially, but we invoke an interactive visualization tool to display the results. Requires pyFLTK (homepage: https://pyfltk.sourceforge.io) and GL_image_display (homepage: https://github.com/dkogan/GL_image_display)
models Camera models representing cameras used to capture the
images. Both intrinsics and extrinsics are used
images The image globs to use for the stereo. If omitted, we
only write out the rectified models. If given, exactly
two image globs must be given-h, --help show this help message and exit
--az-fov-deg AZ_FOV_DEG
The field of view in the azimuth direction, in
degrees. There's no auto-detection at this time, so
this argument is required (unless --already-rectified)
--el-fov-deg EL_FOV_DEG
The field of view in the elevation direction, in
degrees. There's no auto-detection at this time, so
this argument is required (unless --already-rectified)
--az0-deg AZ0_DEG The azimuth center of the rectified images. "0" means
"the horizontal center of the rectified system is the
mean forward direction of the two cameras projected to
lie perpendicular to the baseline". If omitted, we
align the center of the rectified system with the
center of the two cameras' views
--el0-deg EL0_DEG The elevation center of the rectified system. "0"
means "the vertical center of the rectified system
lies along the mean forward direction of the two
cameras" Defaults to 0.
--pixels-per-deg PIXELS_PER_DEG
The resolution of the rectified images. This is either
a whitespace-less, comma-separated list of two values
(az,el) or a single value to be applied to both axes.
If a resolution of >0 is requested, the value is used
as is. If a resolution of <0 is requested, we use this
as a scale factor on the resolution of the first input
image. For instance, to downsample by a factor of 2,
pass -0.5. By default, we use -1 for both axes: the
resolution of the input image at the center of the
rectified system.
--rectification {LENSMODEL_PINHOLE,LENSMODEL_LATLON}
The lens model to use for rectification. Currently two
models are supported: LENSMODEL_LATLON (the default)
and LENSMODEL_PINHOLE. Pinhole stereo works badly for
wide lenses and suffers from varying angular
resolution across the image. LENSMODEL_LATLON
rectification uses a transverse equirectangular
projection, and does not suffer from these effects. It
is thus the recommended model
--already-rectified If given, assume the given models and images already
represent a rectified system. This will be checked,
and the models will be used as-is if the checks pass
--clahe If given, apply CLAHE equalization to the images prior
to the stereo matching. If --already-rectified, we
still apply this equalization, if requested. Requires
--force-grayscale
--force-grayscale If given, convert the images to grayscale prior to
doing anything else with them. By default, read the
images in their default format, and pass those
posibly-color images to all the processing steps.
Required if --clahe
--viz {geometry,stereo}
If given, we visualize either the rectified geometry
or the stereo results. If --viz geometry: we construct
the rectified stereo system, but instead of continuing
with the stereo processing, we render the geometry of
the stereo world; the images are ignored in this mode.
If --viz stereo: we launch an interactive graphical
tool to examine the rectification and stereo matching
results; the Fl_Gl_Image_Widget Python library must be
available
--axis-scale AXIS_SCALE
Used if --viz geometry. Scale for the camera axes. By
default a reasonable default is chosen (see
mrcal.show_geometry() for the logic)
--title TITLE Used if --viz geometry. Title string for the plot
--hardcopy HARDCOPY Used if --viz geometry. Write the output to disk,
instead of making an interactive plot. The output
filename is given in the option
--terminal TERMINAL Used if --viz geometry. The gnuplotlib terminal. The
default is almost always right, so most people don't
need this option
--set SET Used if --viz geometry. Extra 'set' directives to pass
to gnuplotlib. May be given multiple times
--unset UNSET Used if --viz geometry. Extra 'unset' directives to
pass to gnuplotlib. May be given multiple times
--force, -f By default existing files are not overwritten. Pass
--force to overwrite them without complaint
--outdir OUTDIR Directory to write the output into. If omitted, we
user the current directory
--tag TAG String to use in the output filenames. Non-specific
output filenames if omitted
--disparity-range DISPARITY_RANGE DISPARITY_RANGE
The disparity limits to use in the search, in pixels.
Two integers are expected: MIN_DISPARITY
MAX_DISPARITY. Completely arbitrarily, we default to
MIN_DISPARITY=0 and MAX_DISPARITY=100
--valid-intrinsics-region
If given, annotate the image with its valid-intrinsics
region. This will end up in the rectified images, and
make it clear where successful matching shouldn't be
expected
--range-image-limits RANGE_IMAGE_LIMITS RANGE_IMAGE_LIMITS
The nearest,furthest range to encode in the range
image. Defaults to 1,1000, arbitrarily
--stereo-matcher {SGBM,ELAS}
The stereo-matching method. By default we use the
"SGBM" method from OpenCV. libelas isn't always
available, and must be enabled at compile-time by
setting USE_LIBELAS=1 during the build
--sgbm-block-size SGBM_BLOCK_SIZE
A parameter for the OpenCV SGBM matcher. If omitted, 5
is used
--sgbm-p1 SGBM_P1 A parameter for the OpenCV SGBM matcher. If omitted,
the OpenCV default is used
--sgbm-p2 SGBM_P2 A parameter for the OpenCV SGBM matcher. If omitted,
the OpenCV default is used
--sgbm-disp12-max-diff SGBM_DISP12_MAX_DIFF
A parameter for the OpenCV SGBM matcher. If omitted,
the OpenCV default is used
--sgbm-pre-filter-cap SGBM_PRE_FILTER_CAP
A parameter for the OpenCV SGBM matcher. If omitted,
the OpenCV default is used
--sgbm-uniqueness-ratio SGBM_UNIQUENESS_RATIO
A parameter for the OpenCV SGBM matcher. If omitted,
the OpenCV default is used
--sgbm-speckle-window-size SGBM_SPECKLE_WINDOW_SIZE
A parameter for the OpenCV SGBM matcher. If omitted,
the OpenCV default is used
--sgbm-speckle-range SGBM_SPECKLE_RANGE
A parameter for the OpenCV SGBM matcher. If omitted,
the OpenCV default is used
--sgbm-mode {SGBM,HH,HH4,SGBM_3WAY}
A parameter for the OpenCV SGBM matcher. Must be one
of ('SGBM','HH','HH4','SGBM_3WAY'). If omitted, the
OpenCV default (SGBM) is used
--write-point-cloud If given, we write out the point cloud as a .ply file.
Each point is reported in the reference coordinate
system, colored with the nearest-neighbor color of the
camera0 image. This is disabled by default because
this is potentially a very large filehttps://www.github.com/dkogan/mrcal
Dima Kogan, <dima@secretsauce.net>
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
http://www.apache.org/licenses/LICENSE-2.0