Abstract
Purpose: :
The ModelFest data to has been used construct a Standard Spatial Observer model for small target contrast detection. The model has three steps: the contrast image 1) is filtered by a contrast sensitivity function (CSF) including an oblique effect, 2) is windowed by a Gaussian aperture to represent the higher sensitivity of the fovea, 3) is summed over space with a summation exponent near 2.5. The model fit the data from 16 observers for 43 stimuli with an RMS error of 1.0 dB. Since the observer x image interaction error was 0.6 dB, the authors concluded that significant improvement in the model was still possible.
Methods: :
Additional fitting was done with modifications to the model or data.
Results: :
1) Two of the stimuli were excluded: the noise sample and the San Francisco scene. These two stimuli have much more entropy, which strongly reduces detectability. Absent those two stimuli, the RMS drops to 0.7 dB, and the benefit of the oblique effect is negligible.2) On the large Gabors all 16 observers detected the large Gabor at 4 cpd better than at 3 or 5.6 cpd. Nine of the 41 images are 4 cpd Gabors. The hypothesis that these 9 could be fit by the same CSF as the rest was statistically rejected. The next most frequent pattern has only 4 images, so this effect is likely to be an artifact of perceptual learning. With this effect included, the fit of the model is lowered to less than 0.6 dB, the level of the observer by image interaction.3) Previously estimated Gaussian window functions ranged in SD from 0.35 to 0.5 deg, predicting negligible sensitivity in the parafovea. Assuming each cone contributes a constant amount of noise, the contrast sensitivity will be proportional to the square root of the cone density or directly proportional to the estimated Nyquist frequency. Fitting the human cone density measurements of Curcio, et al. (1992 J Comp. Neurol.), we obtain a windowing function with no free parameters that maintains peripheral sensitivity and fits the Modelfest data as well as the Gaussian window.
Conclusions: :
A simpler, more plausible version of the Spatial Standard Observer model accounts for essentially all the predictable variation in the 16 observer by 41 image ModelFest data.
Keywords: detection • computational modeling • learning