Learning attributes of real-world objects by clustering multimodal sensory data

The goal of this work is to propose a framework for learning attributes of real-world objects via a clustering-based approach that aims to reduce the amount of human effort required in the form of labels for object categorization. Due to clustering, with just a single annotation, we can get information about all the objects in a cluster. In the field of robotics, even though studies have focused on the problem of object categorization, the aspect of the amount of workload for a user has not been explored much. However, as the presence of robots has started growing in our daily lives, it is important to reduce the human effort required in labelling for a robot to learn about its environment. Therefore, we propose a hierarchical clustering-based model that can learn the attributes of objects without any prior knowledge about them. It clusters multi-modal sensory data obtained by exploring real-world objects in an unsupervised fashion and then obtains labels for these clusters with the help of a human and uses this information to predict attributes of novel objects.