Abstract
Individuals can give a single color name to a natural object, even though these objects often have color variations across their surfaces. However, it is unclear how observers exploit the information in natural objects to give them a single color name. Since autumn leaves have varying color distributions, we instructed 8 naive observers to assign color names to photographed autumn leaves. Observers viewed high-resolution, 16-bit photographs (constant focus, camera distance, and background conditions with a D65 illuminant) of 275 leaves that ranged from pure "red" to pure "green". Observers indicated whether each leaf appeared "red" or "green" with two buttons. The leaves in each photo were segmented from the background and converted to their corresponding DKL coordinates. Each leaf's isoluminant color distribution mostly resided in the lower right quadrant of the isoluminant plane (the "red-yellow" quadrant). In particular, leaves that appeared "green" had color distributions close to and extended along the "Yellow" direction, with little distribution along the "Green" direction. Various statistics were computed for the isoluminant color distributions (mean, standard deviation, number of "red"/"green"/"yellow" pixels). To find the most informative statistics, linear classifiers were trained on every combination of these statistics, using a "leave-one-out" method: cycle through observers, use the other observers' data for training, and then use the trained classifier to predict the excluded observer's categorizations. For each leaf, the most frequent color name assigned by observers in the training set was taken as "ground truth". On average, the mean color predicts 92% of observer's classifications, explaining most of the variance. Including the standard deviations of each leaf's color distribution along the R-G and Y-V cardinal axes, as well as the number of "red"/"green"/"yellow" pixels, marginally improves the average prediction rate to 93%. Thus, when assigning color names to objects, observers might give little weight to spatial structure.
Meeting abstract presented at VSS 2014