Score to Learn: A Comparative Analysis of Scoring Functions for Active Learning in Robotics.

Accurately detecting objects in unconstrained settings is crucial for robotic agents, such as humanoids, that function in ever-changing environments. Current deep learning based methods achieve remarkable performance on this task on general purpose benchmarks and they are therefore appealing for robotics. However, their high accuracy comes at the price of computationally expensive off-line training and extensive human labeling. These aspects make their adoption in robotics challenging, since they prevent rapid model adaptation and re-training to novel tasks and conditions. Nonetheless, robots, and especially humanoids, being embodied in the surrounding environment, have access to streams of data from their sensors that, even though without supervision, might contain information of the objects of interest. The Weakly-supervised Learning (WSL) framework offers a set of tools to tackle these problems in general-purpose Computer Vision. In this work, we aim at investigating their adoption in the robotics domain which is still at a preliminary stage. We build on previous work, studying the impact of different, so called, scoring functions, which are at the core of WSL methods, on Pascal VOC, a general purpose dataset, and a prototypical robotic setting, i.e. the iCubWorld-Transformations dataset.