16.16. corr_evalΒΆ

The corr_eval program takes as input left and right aligned images and a disparity, as produced by parallel_stereo (Section 16.47), and creates an image of same dimensions as the left image, having a measure of the quality of the disparity at each pixel.

The input disparity can be any of D.tif, B.tif, RD.tif, or F.tif. (Section 19 describes these.)

Two quality metrics are supported, for the moment.

  • Normalized cross-correlation (NCC). For each left aligned image pixel find the normalized cross-correlation between the patch of dimensions given by the kernel size around that pixel and corresponding patch in the right aligned image, as determined by the disparity at that pixel. A higher value means a more reliable disparity.

    Since the disparity is floating-point, bilinear interpolation is used in the right image (unless the flag --round-to-int is set, when the disparity is rounded to int and no interpolation happens). Pixels with no-data values and out-of-range pixels are excluded from the calculation. Note that NCC is not the zero-normalized cross-correlation, so there is no subtraction from each pixel of the mean of all pixels in the patch.

  • Average of standard deviations of left and right matching patches.

The output image has no-data values at pixels where it could not compute the desired metric.

Usage:

corr_eval [options] <L.tif> <R.tif> <Disp.tif> <output prefix>

Example for a disparity created with --stereo-algorithm asp_bm:

corr_eval --prefilter-mode 2 --kernel-size 21 21 --metric ncc \
  run/run-L.tif run/run-R.tif run/run-RD.tif run/run

Example for a disparity created with the asp_mgm algorithm:

corr_eval --prefilter-mode 0 --kernel-size 5 5 --metric ncc \
  run/run-L.tif run/run-R.tif run/run-RD.tif run/run

This will create run/run-ncc.tif.

It appears that always using --prefilter-mode 2 makes this tool produce better results. This is to be investigated further. Consider also experimenting with --kernel-size.

The external stereo algorithms (Section 18) do not use the prefilter mode, and each of them specifies the correlation kernel size in its own way.

See also the somewhat-related image correlator tool which can find the disparity of two images without assuming camera information (Section 16.17).

Command-line options for corr_eval:

--kernel-size <integer integer (default: 21 21)>

The dimensions of image patches. These must be positive odd numbers.

--metric <string (default: ncc)>

The metric to use to evaluate the quality of correlation. Options: ncc, stddev.

--prefilter-mode <integer (default: 0)>

Prefilter mode. This is the same prefilter as in stereo correlation (Section 17) with the asp_bm method. Options: 0 (none), 1 (subtracted mean), 2 (LoG).

--prefilter-kernel-width <float (default: 1.5)>

The diameter of the Gaussian convolution kernel for prefilter modes 1 and 2. A value of 1.5 works well for LoG, and 25 - 30 is suggested for subtracted mean.

--sample-rate <integer (default: 1)>

Compute the quality image only at one out of this many rows and columns, for speed. The output image size does not change. To shrink it (say by 2x), run gdal_translate -r average -outsize 50% 50% in.tif out.tif.

--round-to-int

Round the disparity to integer and skip interpolation when finding the right image patches. This make the program faster by a factor of about 2, without changing significantly the output image.

--threads <integer (default: 0)>

Set the number of threads to use. By default use the number of threads as given in .vwrc, which can be 8 or 16. (The actual number will be printed when this program starts.)

--cache-size-mb <integer (default = 1024)>

Set the system cache size, in MB.

--tile-size <integer (default: 256 256)>

Image tile size used for multi-threaded processing.

--no-bigtiff

Tell GDAL to not create bigtiffs.

--tif-compress <None|LZW|Deflate|Packbits (default: LZW)>

TIFF compression method.

-v, --version

Display the version of software.

-h, --help

Display this help message.