Comments on "Recognition of faces in the presence of two-dimensional sinusoidal masks" by Tieger and Ganz

Identification of band-pass filtered letters and faces by human and ideal observers

To better understand how the visual system makes use of information across spatial scales when identifying different kinds of complex patterns, we measured human and ideal contrast identification thresholds to estimate identification efficiency for 1- and 2-octave wide band-pass filtered letters and faces embedded in 2-D dynamic Gaussian noise. Varying stimulus center frequency from 1 to 70 c/object had different effects on letter and face identification efficiency. In the 2-octave conditions, identification efficiencies decreased by 0.25-0.5 log units for letters and 0.5-1.2 log units for faces as center frequency increased from 6.2 to 49.5 c/object, but only letters were identifiable at center frequencies below 6.2 c/object. In the 1-octave conditions, letter identification efficiencies increased by about 0.5 log units as center frequency increased from 1.1 to 2.2 c/object, and were nearly constant from 2.2 to 35 c/object. Letters were unidentifiable by human observers at 70 c/object. Surprisingly, face identification was impossible for human observers at all center frequencies except 8.8 c/object for one observer, and 8.8 and 17.5 c/object for a second observer. Ideal observer thresholds were obtained for both letters and faces in all conditions, so information was always available to perform the task. Thus, the failure to identify faces reflects constraints on visual processing rather than a lack of stimulus information. Selective spatial sampling may account for some of the differences between letter and face identification efficiencies.

Perception of visible speech: influence of spatial quantization

Visible speech reading was studied to determine which features are functional and to test several models of pattern recognition. Nine test syllables differing in their initial consonant were presented in intact form or under various levels of spatial quantization. Performance decreased in increasing quantization but remained relatively good at moderate levels of degradation. Different models were tested against the confusion matrices. Six features were identified as functional in distinguishing among the nine consonant--vowel syllables. These features were used as sources of information in a fuzzy-logical model of perception and an additive model. The fuzzy-logical model provided a significantly better description of the confusion matrices, showing that speech reading is analogous to other domains of pattern recognition such as face recognition and facial-affect perception.