Recognizing Unknown Objects for Open-Set 3D Object Detection

While most studies in computer vision focus on closed set recognition, where the number of possible categories is fixed and known a priori, this study provides new insights into open set recognition. Open set problems assume that there is an unlimited number of categories, and most of them are unknown. This work contributes to our understanding of the unknown and how it can be applied in the object detection setting. To date, research on open set recognition has been focused chiefly on images or sequences of images. The novelty of our work lies in an adaption of 3D object detection to an open set setting. We analyse the detections from the popular instance segmentation framework, discuss the object detection performance on different examples, and show how unknown objects can be detected. We describe our approach to build a 3D open set object detection system, implement it in the simulation, and provide tools for evaluation. Moreover, this work demonstrates the method of learning novel classes without manual data labelling. We hope that the results of our work will bring robots closer to the ability to know that they do not know, and benefit from this finding.