Minerva scene analysis benchmark

Abstract

The analysis of natural scenes is an important research area. Scene analysis research provides the foundation for the development of autonomous systems whose vision sensors provide important information about the surrounding environment. In this paper we introduce the Minerva scene analysis benchmark to the vision community and provide preliminary results on this data. The scene analysis benchmark contains 448 natural images in both colour and greyscale format. The images contain 8 natural objects including sky, brick, clouds, pebbles, road, trees, grass and leaves. The benchmark is intended to facilitate further research into scene analysis and to encourage the development of tools and techniques that work on natural object recognition. The results reported here have used four image segmentation techniques including fuzzy c-means clustering, histogram based thresholding, region growing, and split and merge. Following segmentation, texture features for object classification have been generated using five different texture analysis methods including autocorrelation, co-occurrence matrices, edge frequency, Law's, and run length. These results can be taken as a preliminary baseline on this benchmark.