Deep Learning Performance in the Presence of Significant Occlusions - An Intelligent Household Refrigerator Case

摘要

Real-world environments, inhabited by people, still pose significant challenges to deep learning methods. Object occlusion is one of such problems. Humans deal with the occlusion in a complex way, by changing the viewpoint and using hands to manipulate the scene. However, not all robotic systems can do that due to cost or design constraints. The question we address in this paper is, how well modern object detection methods work on a model case of an intelligent household refrigerator, where numerous occlusions occur. To motivate our research, we actually performed a worldwide survey of refrigerator occupancy to realistically judge the extent of the problem, but the results could be generalized to any unstructured storage environment where people are in charge. The survey results enabled us to generate a dataset of photo-realistic renderings of a typical refrigerator interior, where the object identity, location, and the degree of the refrigerator occupancy are all readily available. Our results are represented as the Average Precision depending on a refrigerator occupancy for two well known deep models.