查看文章

Object detection is a robot perception task that requires classifying objects in the scene into one of many predefined categories, as well as localizing these objects through estimating their tightest fitting bounding boxes. Modern deep object detectors produce point estimates of object categories and bounding boxes in the scene, making it difficult for subsequent components of a robot system from quantifying the trustworthiness of the objects detected in the scene. A solution to this problem is to construct probabilistic object detectors, models that can predict probability distributions rather than point estimates of object categories and bounding boxes in the scene. In this dissertation, we present several contributions to deterministic and probabilistic object detectors. We first present AVOD, one of the first deterministic object detectors to efficiently solve the 3D object detection problem. We then present BayesOD, a …