Improving color homogeneity measure in superpixel segmentation assessment
ResumoThe quality of a superpixel segmentation may consider accuracy in delineation, shape compactness, and color homogeneity. Among several existing measures, Explained Variation (EV) and Intra-cluster Variation (IV) seem to be the only ones focusing on color homogeneity. However, EV ignores color differences inside superpixels while IV reduces penalization by averaging those differences. This work proposes a superpixel color descriptor to measure color homogeneity when comparing superpixel algorithms. Our RGB-cube Bucket Descriptor (RBD) is a compact representation of the most relevant colors in each superpixel. Color homogeneity is measured based on differences between pixel color and its closest color in RBD and the color differences inside RBD. We call it Similarity between Image and Reconstruction from Superpixels (SIRS) since, substituting each pixel color by its closest color in RBD, one obtains an image reconstruction. A high-quality superpixel segmentation (c) (d) should then present a reconstruction similar to the original image. Experiments on three datasets show that SIRS can better distinguish segmentation algorithms according to color homogeneity than EV (the most popular measure). The results also show that SIRS is more robust to slight color variations due to luminosity than EV.
Palavras-chave: Image segmentation, Image color analysis, Shape, Focusing, Birds, Robustness, Proposals
BARCELOS, Isabela B.; BELÉM, Felipe De C.; JOÃO, Leonardo De M.; FALCÃO, Alexandre X.; GUIMARÃES, Silvio J. F.. Improving color homogeneity measure in superpixel segmentation assessment. In: CONFERENCE ON GRAPHICS, PATTERNS AND IMAGES (SIBGRAPI), 35. , 2022, Natal/RN. Anais [...]. Porto Alegre: Sociedade Brasileira de Computação, 2022 .