On The Fly Object Category Learning

Abstract

Image search methods (for example, Search On Google), rely on text features. This results in uncertain and noisy outcomes. Robots are getting more intelligent, efficient, and capable of completing more challenging tasks, as Artificial Intelligence (AI) advances. Recent advancements have revolutionized the field of artificial intelligence. Rather than performing pre-programmed tasks, robots are now learning new things and becoming more autonomous, as a result. However, in most cases, robots require some level of human assistance to learn something. Recognizing and classifying everyday objects is a critical skill for a service robot. In this paper, we implemented a fully autonomous object category learning system for service robots, in which the robot learns object categories using internet resources. Our goal is to train a model and obtain data against a query from another dataset. Previous work demonstrates that by providing a text summary, web data is fetched and categories of objects can be learned by model and features can be used to fetch the same categories from other datasets to rank text-based search results. A practical and effective option to massively improve the user interface, for another dataset, is to use Web Search. The robot retrieves images of the object from the internet and uses them to generate training data for learning classifiers. The classification performance is compared to a benchmark dataset. On 101 object categories from the benchmark dataset, the system performed well, with 98.21 percent average accuracy and 96.44 percent average precision, and showed promising results in each scenario. There are currently ongoing research projects dealing with object category learning for robots using internet images. For that class, we use two measures to evaluate the on-the-fly learning model: Accuracy and Precision. Keywords —– On The Fly Learning For Retrieval, Run Time Learning, Category Learning For Retrieval, Instance and Model Based Learning.