Scoring functions evaluation for active learning in humanoid robotics

Abstract: Object detection is an essential ability for humanoid robots. State-of-the art deep learning models for object detection reached remarkable results on general-purpose datasets. However, these methods are typically based on the supervised learning framework, requiring large amounts of data and long computational time to be trained. For these reasons, they are not suited for robots that are required to interact in unconstrained environments, since the annotation process of training data is typically expensive and time consuming. Active learning techniques address these issues by choosing for labeling only the most informative images of a given unlabeled dataset. This master thesis project presents an extensive empirical analysis of the so called "scoringfunctions" [1] (also referred as "queries selectors"), which define the concept of informativeness of images, providing the possibility to compare them, and therefore, to consider the most useful ones for training. The self-supervision sample mining (SSM) [2] pipeline is used as baseline and starting point of the proposed empirical analysis since the considered learning protocol has proven to be particularly useful for robotic applications [3]. The considered methods have been evaluated on the PASCAL VOC [4], a standard computer vision benchmark, and on the iCubWorld Transformations [5], a challenging robotics dataset. The presented results on both datasets allow drawing interesting conclusions. Specifically, the analysis shows that efficient solutions can be identified in the robotic scenario, reaching high precision values with few images added to the initial training set, and in a fraction of the selection time.