Semantics Driven Resampling of the OSA-UCS

Abstract

In this paper, we propose a resampling of the OSA-UCS. Following the same sampling criterion that used to define the 424 specimens of the OSA-UCS set, we enlarged such an ensemble by adding 590 samples located in the outer region of the original volume. This allows to overcome the bottleneck in the use of the original color basis for computer vision applications due to lack of saturated colors. The outcomes of a color categorization experiment performed on the extended basis were used to train a discrete color naming model that we have recently proposed. The model was validated through the analysis of its performance for segmenting natural images. Results show that the extended basis removes the inability of the model to deal with saturated colors which significantly improves segmentation results and makes the extended bases exploitable for computer vision applications.