Using the example plugins

MRF transforms

MRF transforms require a delineated subject (or list of subjects) to transform, and two sets of delineated images, one for the "start" group (should match the subject) and one for the "target" group. MRF transforms work better with the first derivative filters than with the default (Gabor) filters, these are available in the extras.zip file from the download page:

• ​http://users.aber.ac.uk/bpt/jpsychomorph/extra.zip -- some additional useful data files

These can be loaded via the main window menu Average->Texture Options -> Load Filters.

When you select the MRF Transform option you are prompted for the two sets of files. You are also asked for the coefficients of a quadratic equation that governs how the smoothing of local feature histograms should vary with spatial scale. The values "1 0 0" (a constant value of 1) will lead to a crisp transform appearance, but the face will likely be made largely from a single individual in the output set. Using a linear equation (e.g. values "1 0.5 0") will result in a slightly blurrier transform, but retaining more of the original subject's identity. Using a quadratic equation (e.g. values "1 0.5 0.5") will increase this effect further.

Another option asks you if you would like to use a Gaussian approximation to the local conditional histograms. Answering "Yes" seems to give almost identical results to using the full histograms and is considerably faster.

Colour calibration

The colour calibration assumes that you have a colour chart located in approximately the same place in each image. Select the colour chart using the rectangle in the delineator window (View -> Display rectangle, then click and drag). The box should be a reasonably tight fit to the chart. When you select colour calibrate, the algorithm samples circles regularly spaced on a 4 (wide) by 6 (high) grid, and iteratively tries to segment pixels of each colour and align the grid to the segmented colour patches. This seems to improve the alignment (e.g. for small rotations). The mean colours from each patch are then compared with the known colours to build a non-linear model of the colour mapping and apply it to the image. In batch mode you are given the option to use the calibration from the first image for all images, or to update the sampled colours (by estimating the grid position, scale and orientation) for each image.

The known colours are supplied in a file, starting with the whitepoint for the measured values (e.g. 65 for the d65 illuminant), followed by each patches colour. The colours are ordered from top to bottom and left to right. An example is given below, in this example the first 6 values represent the grey values from white to black, which is the column on the (viewers) left of the chart.

65
96.07	-0.74	1.73
80.39	0.09	-0.06
66.34	-0.16	-0.63
53.88	-0.03	-0.03
40.89	0.06	-0.26
28.99	-0.1	-0.87
33.12	11.9	-42.25
54.94	-34.45	33.18
40.81	47.21	22.18
78.08	3.8	76.12
51.7	42.41	-16.02
54.02	-27.81	-22.42
60.65	33.43	52.65
43.78	7.58	-40.38
49.68	41.5	13.58
31.61	18.65	-21.02
71.26	-19.92	56.62
68.85	18.96	62.97
38.25	12.9	13.1
64.48	13.48	17.16
51.41	-3.98	-20.47
43.07	-11.07	20.64
55.95	6.89	-23.33
71.26	-29.71	2.1

Symmetrising

In order to symmetrise an image you need to have the image delineated with points and lines, and you need to specify the mirrored counterpart of each point. The points for a particular Template type can be labelled interactively from the Delineator window (Delineate -> Label symmetry). Each point is highlighted and you need to click on the corresponding mirrored point. The results are saved in a .sym file. The extra.zip file (see link above) contains an example .sym file for the standard template configuration.