Abstract
Detecting other humans is a basic visual function that occurs under a wide range of variability in appearances, in particular those resulting from distance, normal defocus, or impaired acuity. Our goals were to measure resolution limits of human body detection and to investigate whether detection is primarily determined by the effective number of spatial samples. Participants (N=30) were shown natural, color images of human bodies or background scenes and asked to identify whether each image contained a body in a yes-no paradigm. We selected 176 body images and 176 background scene stimuli. Body images were approximately matched for body size, and background images were matched with body images based on scene categories. There were three degradation conditions in which the amount of spatial information was progressively reduced: 1) Size reduced – images were downsampled from 140x205 pixels to between 21x30 and 2x3 pixels, producing 10 images with retinal sizes ranging from 40x57 to 4x6 minutes of arc when viewed from 50 centimeters; 2) Block-averaged – the size-reduced images were upsampled to 13x19 degrees; 3) Low-pass filtered – the block-averaged images were spatially filtered to suppress frequencies above half the Nyquist rate. Images were displayed for up to five seconds per trial, with participants instructed to emphasize accuracy rather than speed. For each participant, an equal number of body and background images were shown at each resolution, and each image was displayed three times; once using each image degradation type. We found that the effective number of samples required to achieve a sensitivity of d'=1 was 12x18, 12x18 and 9x13 spatial samples for conditions 1, 2, and 3, respectively. While spatial filtering (condition 3) improved performance, solely increasing retinal size of the image (condition 2) had no effect on sensitivity.
Meeting abstract presented at VSS 2018