Abstract:

This paper is organized around the problems of classifying objects depicted in photos. The main issue that is being addressed is determining how many images are necessary to solve such a problem. Naturally, we will consider using the least number of photos in order to minimize the effort required to create the data set. The study case presented in this article is based on a number of low difficulty classification tasks, distinguishing between two and five classes. For further constraining, the samples in the data set pertaining each class (about 50, each) are images of a single object, varying only in pose, perspective, background and viewing conditions. This paper only considers aspects arisen from the study of the behavior of a convolutional network with predefined structure (thus, omitting, for now, the treatment of the dilemma of network design vs. the quality of images). The problems arising from such context will be presented, alongside a method for assessing the level of relevance of the data set for a certain task. The results and conclusions are obtained in an empirical manner, based on 15 classes organised in 9 classification tasks.