ePose: Let's Make EfficientPose More Generally Applicable

TL;DR

This paper introduces ePose, an enhanced version of EfficientPose that can infer object sizes and simplifies data collection and loss calculations, aiming for broader applicability in 3D object detection.

Contribution

ePose extends EfficientPose by enabling size inference and streamlining data and loss processes for improved 3D detection.

Findings

Evaluated on Linemod and Occlusion 1-class datasets

Demonstrated improved efficiency and applicability

Discussed potential use with NuScenes and KITTI datasets

Abstract

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.