Skip to content

Model Parameters

Jump to bottom
Heiko Schütt edited this page Jul 16, 2019 · 2 revisions

There are two early-vision-model functions, which differ in whether we use a foveal window (early_vision_model) or not (early_vision_model_NoFovea). To run these functions requires a parameter set and a few inputs explained on this page:

input parameters

[outN,imageNoise] = early_vision_model_NoFovea(imageBitmap,degSize,timecourse,pars,sizePx,V1Mode,csfselector)

Inputs are:

  • imageBitmap: the image to be processed, should be in cd/m^2
  • degSize: size of the image in degrees of visual angle
  • timecourse: a number or a vector to specify the presentation duration of the stimulus. For vectors, currently only the length of the vector is used. Originally this was intended to hold the millisecond resolution time course of the stimulus.
  • pars: a vector of parameters for the model as described below
  • sizePx: A size of the image in pixels. If set the image is first downscaled to this resolution
  • V1Mode: select normalization pool structure (1 = mean over whole image, 2 = only same pixel)
  • csfselector: csf shape used; usually set to match V1Mode

parameter values

You find the parameter values for the model in the subfolder pars. You can either load them manually or load them with the function getPars. In either case you will have to specify the time condition and the V1Mode, i.e. how long the stimulus was presented and whether we shall include only the given pixel or the whole image into the normalization pool.

The pars vector contains the following entries:

  • noiseConst = pars(1): constant noise variance
  • noiseFactor= pars(2): linear factor for noise
  • CNaka = pars(3): Naka rushton constant
  • ExNaka = pars(4)+pars(5): Naka rushton Exponent top
  • ExNakaNorm = pars(4): Naka rushton Exponent below
  • bw(1) = pars(6): bandwidth in frequency (std of log-Gabor in octaves)
  • bw(2) = pars(7): bandwidth in orientation (std log-Gabor in radiants)
  • nFreq = pars(8): number of frequency bands
  • nOrient = pars(9): number of orientations
  • poolSize = pars(10): meaning depends on type, spatial size of normalization pool
  • poolbw = pars(11): bandwidth of the normalization pool (Octaves in frequency)
  • minF = pars(12): lowest frequency band
  • maxF = pars(13): highest frequency band
  • poolO = pars(14): bandwidth of the normalization pool in orientation

Models used in the article

In the article, we used the early_vision_model function with V1Mode = 2 for the main text. V1Mode is the mean normalisation version we mention in the text and appendix figures. In applications the _NoFovea version might be helpful though, as the cut out is probably too strong for most applications, where one might be interested in modeling fovea quality processing everywhere instead.

Clone this wiki locally