Abstract
Introduction
Form vision is significantly impaired in the periphery. It has been proposed that this is due to the learned statistics in the periphery being confounded by saccadic eye movements.
We use sparse coding to capture natural image statistics. We train one basis with the confounding effects of saccades and one without. We then compare the encoding performance of these two bases to see if they can replicate the deficiencies of the periphery, particularly crowding.
Methods
We produced videos that replicate the input to the entire crowding zone at an eccentricity of 5deg during a saccade. The movies last 66ms, with a saccade beginning between 40 and 60ms, corresponding to the time we hypothesize that the periphery learns image statistics under the spotlight of attention. We trained one basis on these videos and another basis on still natural image movies of the same duration.
We then used these basis sets to encode artificial stimuli. The stimuli consisted of a white target in the center of a gray background with added white noise. We trained a linear SVM to classify the encoded stimuli. We then used this SVM to classify similarly encoded movies with the presence of a flanker.
Results
We found that there was a significant crowding effect with the saccadic basis, that is, the further away the flanker was from the target the better the classification performance. We found that the still basis outperformed the saccadic basis; however there was evidence of crowding with the still basis as well. We further found that the saccadic basis preformed worse in the direction of the saccade than in the tangential direction, reproducing another characteristic of crowding.
Conclusion
Sparse coding provides a viable method to study the interaction of image enocoding and crowding in peripheral vision.
Meeting abstract presented at VSS 2013