Paper ID: 2111.15114

ePose: Let's Make EfficientPose More Generally Applicable

Austin Lally, Robert Bain, Mazen Alotaibi

EfficientPose is an impressive 3D object detection model. It has been demonstrated to be quick, scalable, and accurate, especially when considering that it uses only RGB inputs. In this paper we try to improve on EfficientPose by giving it the ability to infer an object's size, and by simplifying both the data collection and loss calculations. We evaluated ePose using the Linemod dataset and a new subset of it called "Occlusion 1-class". We also outline our current progress and thoughts about using ePose with the NuScenes and the 2017 KITTI 3D Object Detection datasets. The source code is available at https://github.com/tbd-clip/EfficientPose.

Submitted: Nov 30, 2021