Patterns in Nature and Human Visual Perception by Ann Hermundstad

Explore patterns in nature and human visual perception in this hour-long lecture by Ann Hermundstad from the International Centre for Theoretical Sciences. Delve into the complex correlational structure of natural stimuli and investigate which correlations humans are most sensitive to. Learn about local binary patterns, psychophysical experiments using synthetic textures, and how the variability of correlations in natural scenes predicts human sensitivity. Examine the relationship between image statistics and human perception, and discover how local correlations can distinguish objects and boundaries. Gain insights into channel coding problems and the broader class of informative gliders that produce salient patterns. Conclude with a thought-provoking question about potential blindness to informative structures in other imaging modalities.

Syllabus

INTERNATIONAL CENTRE for
Ann M Hermundstad
Putters in nature and
Patterns in nature and human visual perception
natural stimuli involve many patterns of light that have complex correlational structure...
to what correlations should we be most sensitive?
we choose to work in a space of local binary patterns
binary texture space is fully-described up to fourth-order by 10 independent coordinates
we reprocess an ensemble of natural images patches
we reprocess an ensemble of natural images patches and extract correlations from each patch
the ensemble of patches yield a multidimensional probability distribution of image statistics
we can extract correlations from or put correlations into visual textures
psychophysical experiments use synthetic binary textures in a segmentation task
task: identify the location of correlated texture
...or in pairwise combinations
threshold values along cardinal and oblique directions define an discrimination contour
coordinates can be tuned individually ...
we extract correlations from images, and we measure human sensitivity to the same correlations
we predict that the variation in image statistics matches human sensitivity to the same statistics
natural scenes are most variable in second-, then fourth-, then third-order correlations
the variability of correlations in natural scenes predicts human sensitivity to the same correlations
covariance in natural images differs for symmetric versus asymmetric combinations of statistics
shape and orientation of inverse covariances predict human discrimination contours
what we've learned so far...
consider a channel coding problem:
do local correlations distinguish objects and boundaries?
we can separately estimates statistics within foregrounds and backgrounds
the square glider belongs to a broader class of informative gliders that produce salient patterns
are we blind to informative structure in other imaging modalities?
thank you!
[Demo]