Item

Abstract

Apart from shape, color is the most visually prominent and perceivable feature of a flower. To use color as a feature for fine-grained plant species classification based on flower images, its descriptor has to be discriminative, compact, and robust against photometric variations. Therefore, we studied state-of-the-art color description methods and evaluated their discriminative power in an image classification pipeline. Experiments have been performed on three flower image datasets possessing large photometric and geometric varieties. We found that implicit photometric invariance by pooling 11 basic colors from patches around local features allows for robust color description outperforming explicitly photometric invariant descriptors in most cases.