Computing Image Descriptors from Annotations Acquired from External Tools

Abstract

Visual descriptors are widely used in several recognition and classification tasks in robotics. The main challenge for these tasks is to find a descriptor that could represent the image content without losing representative information of the image. Nowadays, there exists a wide range of visual descriptors computed with computer vision techniques and different pooling strategies. This paper proposes a novel way for building image descriptors using an external tool, namely: Clarifai. This is a remote web tool that allows to automatically describe an input image using semantic tags, and these tags are used to generate our descriptor. The descriptor generation procedure has been tested in the ViDRILO dataset, where it has been compared and merged with some well-known descriptors. Moreover, subset variable selection techniques have been evaluated. The experimental results show that our descriptor is competitive in classification tasks with the results obtained with other kind of descriptors.