pamhomography - stretch/shrink a quadrilateral region of an image to another quadrilateral shape
pamhomography [-from=coords] [-to=coords] [-mapfile=map_file] [-view=coords] [-fill=color] [-verbose [pam_file]
You can abbreviate any option to its shortest unique prefix. You can use
two hyphens instead of one to delimit an option. You can separate an option
from its value with whitespace instead of =.
This program is part of Netpbm.
pamhomography stretches and shrinks an arbitrary quadrilateral portion of an input image you specify (not necessarily rectangular), into a new quadrilateral shape you specify, producing a new image.
You can do any affine image transformation: translation, reflection, scaling, rotation, and shearing/skewing. However, pamhomography additionally can do bilinear transforms, which means it can warp any quadrilateral to any other quadrilateral, even when this mapping cannot be described using a single set of linear equations. This can be useful, for example, for creating perspective views of rectangular images or for reverse-mapping a perspective view back to a rectangular projection.
In addition to the options common to all programs based on libnetpbm (most notably -quiet, see Common Options), pamhomography recognizes the following command line options:
This defines the source quadrilateral. coords is a list of four integer-valued (x, y) coordinates. If you do not specify the source with either -from or -mapfile, the source quadrilateral is the entire input image.
This defines the target quadrilateral. coords is a list of four integer-valued (x, y) coordinates. If you do not specify the target with either -to or -mapfile, the target quadrilateral is the same as the entire input image.
This names a text file that describes the mapping from the source to the target quadrilateral. The file map_file must contain either eight integer-valued (x, y) coordinates, being the four source coordinates followed by the corresponding four target coordinates, or only four (x, y) coordinates, being only the four target coordinates. In the latter case, the source quadrilateral is taken to be the four corners of the input image in clockwise order, starting from the upper left.
Anything you specify with -to or -from overrides what is in this file.
This defines the target view. coords is a list of two integer-valued (x, y) coordinates: the upper left and lower right boundaries, respectively, of the pixels that will be visible in the output image. If -view is not specified, the target view will fit precisely the target quadrilateral.
This is the color with which the program fills all pixels that lie outside of the target quadrilateral. Specify the color as described for the argument of the pnm_parsecolor() library routine.
The default is black, and for images with a transparency plane, transparent.
Cooordinates should normally be specified in clockwise order. The syntax is fairly flexible: all characters other than the plus sign, minus sign, and digits are treated as separators. Although coordinates need to be integers, they may lie outside the image's boundary.
pamhomography's only parameter, pam_file, is the name of the file containing the input image. If you don't specify pam_file, the image comes from Standard Input.
The output image uses the same Netpbm format as the input image.
Simple transformations are best handled by other Netpbm programs, such as those listed in the "SEE ALSO" section below. Use pamhomography for more sophisticated transformations such as perspective adjustments, rotations around an arbitrary point in the image, extraction of non-rectangular quadrilaterals, shearings by coordinates rather than by angle, and, in general, all transformations that are most easily expressed as mapping four points in one image to four points in another image.
The following examples use the park_row.ppm test image, which is a photograph of New York City's Park Row Building, scaled to 441×640, converted to a PPM file, and redistributed under the terms of the GFDL.
The first example showcases the real power of bilinear transformations. Assuming park_row_rect.map has the following contents:
(147, 51) (316, 105) (402, 595) (92, 560)
(0, 0) (440, 0) (440, 639) (0, 639)then
pamhomography -mapfile park_row_rect.map park_row.ppm > park_row_rect.ppmprojects the building's facade from a perspective view to a rectilinear front-on view. Remember that pamhomography ignores the parentheses and commas used in park_row_rect.map; they merely make the file more human-readable. We equivalently could have written
147 51 316 105 402 595 92 560 0 0 440 0 440 639 0 639or any of myriad other variations.
pamhomography can warp the image to a trapezoid to make it look like it's leaning backwards in 3-D:
pamhomography -to "50,0 390,0 440,200 0,200" park_row.ppm > park_row_trap.ppmAs a very simple example,
pamhomography -to "440,0 0,0 0,639 440,639" park_row.ppm > park_row_flip.ppmflips the image left-to-right. Note that in this case the target quadrilateral's coordinates are listed in counterclockwise order because that represents the correspondence between points (0, 0) ↔ (440, 0) and (0, 639) ↔ (639, 0).
Scaling is also straightforward. The following command scales down the image from 441×640 to 341×540:
pamhomography -to "0,0 340,0 340,539 0,539" park_row.ppm > park_row_small.ppmLet's add 100 pixels of tan border to the above. We use -view and -fill to accomplish that task:
pamhomography -to "0,0 340,0 340,539 0,539" -view "-100,-100 440,639" -fill tan park_row.ppm > park_row_small_border.ppmWe can add a border without having to scale the image:
pamhomography -view "-100,-100 540,739" -fill tan park_row.ppm > park_row_border.ppmThe -view option can also be used to extract a rectangle out of an image, discarding the rest of the image:
pamhomography -view "130,10 205,80" park_row.ppm > park_row_cut.ppmSpecifying the same set of coordinates to -from and -to has the same effect but also allows you to extract non-rectangular quadrilaterals from an image:
pamhomography -from "185,300 310,325 320,425 180,405" -to "185,300 310,325 320,425 180,405" park_row.ppm > park_row_cut_2.ppmRotation is doable but takes some effort. The challenge is that you need to compute the rotated coordinates yourself. The matrix expression to rotate points \((x_1, y_1)\) \((x_2, y_2)\), \((x_3, y_3)\), and \((x_4, y_4)\) clockwise by \(\theta\) degrees around point \((c_x, c_y)\) is
\[ \begin{bmatrix} 1 & 0 & c_x \\ 0 & 1 & c_y \\ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} \cos \theta & -\sin \theta & 0 \\ \sin \theta & \cos \theta & 0 \\ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} 1 & 0 & -c_x \\ 0 & 1 & -c_y \\ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_1 & x_2 & x_3 & x_4 \\ y_1 & y_2 & y_3 & y_4 \\ 1 & 1 & 1 & 1 \end{bmatrix} \quad. \]
For example, to rotate park_row.ppm 30° clockwise around (220, 320) you would compute
\[ \begin{bmatrix} 1 & 0 & 220 \\ 0 & 1 & 320 \\ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} \cos 30^{\circ} & -\sin 30^{\circ} & 0 \\ \sin 30^{\circ} & \cos 30^{\circ} & 0 \\ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} 1 & 0 & -220 \\ 0 & 1 & -320 \\ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} 0 & 440 & 440 & 0 \\ 0 & 0 & 639 & 639 \\ 1 & 1 & 1 & 1 \end{bmatrix} = \begin{bmatrix} 189.4744 & 570.5256 & 251.0256 & -130.0256 \\ -67.1281 & 152.8719 & 706.2621 & 486.2621 \\ 1.0000 & 1.0000 & 1.0000 & 1.0000 \end{bmatrix} \quad, \]
round these coordinates to integers, transpose the matrix, and produce the following map file, park_row_rot30.map:
189 -67
571 153
251 706
-130 486(These are the "to" coordinates; we use the default, full-image "from" coordinates.) The mapping then works as in all of the preceding examples:
pamhomography -mapfile park_row_rot30.map park_row.ppm > park_row_rot30.ppmpamhomography was new in Netpbm 10.94 (March 2021).
Copyright © 2020 Scott Pakin, scott+pbm@pakin.org