Reproducing and Comparing Results: Retina-V1 model of detectability across the visual field
Author: James Capps
Summary
This project is based on a recent paper by Chris Bradley, Jared Abrams, and Wilson S. Geisler, that studies the detectability across the visual field using a Retina-V1 model. The model is designed to predict "the detectablility of targets at arbitrary locations in the visual field, in arbitrary gray-scale backgrounds, and under photopic viewing conditions." The model presented in the paper was coded with MATLAB. For this project, I attempted to reproduce the results of the model by rewriting the front-end part of the model into ISETBIO, compare my findings to those of the paper, and explain any discrepencies.
Background
In 2014, Chris Bradley, Jared Abrams, and Wilson S. Geisler designed a model for "predicting the detectability of targets at arbitrary locations in the visual field, in arbitrary gray scale backgrounds, and under photopic viewing conditions." They developed this model as a way to streamline existing knowledge of the subject into a practical method of rapidly determining target locations on the retina. The model can be divided into two main components: the retinal and cortical components. The first component focuses on the functionality of the human retina and visual response of the midget ganglion cells, taking into account human optics and light adaptation. The cortical component reproduces the spatial pattern masking effect of the background and pools the responses to the primary visual cortex.
Goal
This project seeks to implement the front-end of the Retina-V1 model (specifically the Optics and Light Adaptation sections) with ISETBIO and to compare its functionality to that of the original MATLAB code of the original model.
Retinal Model
Optics
The retinal images of the target and background are computed by convolving the target and background with an optical point spread function. Bradley, Abrams, and Geisler used the point spread function (psf) recorded in Navarro et al. (1993) of an average human psf in the fovea. Then they took the Fourier transform and multiplied it by the modulation transfer function (mtf) also reported by Navarro et al.
Light Adaptation
Light adaptation is the means by which the eye maintains light discrimination and sensitivity while keeping neuron responses in the retina within their dynamic range. It is calculated as a multiplicative luminance gain control. Bradley et al. find the local luminance by using a 2-D gaussian centered on retinal location. They then take the inverse of this average local luminance to find the luminance gain.
ISETBIO
Optics
In order to see how comparable the MATLAB code formed by Bradley et al. is to ISETBIO, a human optics condition must be created. This comparison can be simulated by creating an average human modulation transfer function (mtf) with ISETBIO and testing it against the mtf provided by Bradley et al. Comparing human optical transfer functions (otf) at different wavelengths to form the mtf, the wavelength that seems to most closely match that of the mtf in the paper is about 540 nm. This graph is shown to the right.
Light Adaptation
Simulating light adaptation, like the optics portion, requires human conditions. The Bradley paper manages this by interpolating a series of multi-resolution stacks. The ISETBIO code creates a human optical image of a scene, as well as a human sensor to view this scene, and measures the light adaptation gain of the cones in the sensor. The local luminance gain can then be obtained from this value.
Comparing the MATLAB results to those of ISETBIO, the numbers were close but not identical. MATLAB produced a luminance gain of about 0.0078, while ISETBIO's luminance gain was around 0.0100. This difference could be a result of the different backgrounds used in each.
Conclusion
Aside from a few small differences, ISETBIO seems to have the ability to replicate the retinal component of the MATLAB code of this paper. ISETBIO is convenient and easy to use, and it has the robustness to accurately represent the functionality of optics and the human eye.
References
Bradley C., Abrams J., Geisler W. "Retina-V1 model of detectability across the visual field." Journal of Vision (2014) http://www.journalofvision.org/content/14/12/22, doi:10.1167/14.12.22.
Code
Original MATLAB code: http://natural-scenes.cps.utexas.edu/retina_V1_model/
ISETBIO: https://github.com/isetbio/isetbio
Compare Code: Media:compare_bradleyJOV2014_ISETBIO.zip