This work investigates the uncertainty in 3D pose estimation of DenseFusion, a Deep Learning-based 3D Object Pose Estimation (3DOPE) framework. The accuracy and precision of object spatial localization is evaluated on the benchmark Yale Carnegie Berkeley-Video Dataset. The Average Euclidean Distance metric is employed as a figure of merit for the quantitative data analysis. Moreover, the Average Rotational and Translational Distances among the estimated and the ground-truth poses are analyzed. Results gathered on the YCB- Video Dataset are validated on a newly collected 3DOPE dataset, the ISS Dataset. The latter is populated by harnessing a 3D scanner, an RGB-D camera, and an anthropomorphic robot. Outcomes highlight an Average Euclidean Distance value of 1.8 mm on the best-processed item in the YCB-Video Dataset. Similarly, an Average Euclidean Distance value of 1.9 mm is achieved on the ISS Dataset. Eventually, the measurement uncertainty of the six components of the 3D poses is inspected. The latter assessment is performed on the ISS Dataset by relying on the ISO GUM [1].

3DOPE-DL: Accuracy Evaluation of a Deep Learning Framework for 3D Object Pose Estimation

Fabris, Davide Maria;Sala, Remo;Marco, Tarabini
2020-01-01

Abstract

This work investigates the uncertainty in 3D pose estimation of DenseFusion, a Deep Learning-based 3D Object Pose Estimation (3DOPE) framework. The accuracy and precision of object spatial localization is evaluated on the benchmark Yale Carnegie Berkeley-Video Dataset. The Average Euclidean Distance metric is employed as a figure of merit for the quantitative data analysis. Moreover, the Average Rotational and Translational Distances among the estimated and the ground-truth poses are analyzed. Results gathered on the YCB- Video Dataset are validated on a newly collected 3DOPE dataset, the ISS Dataset. The latter is populated by harnessing a 3D scanner, an RGB-D camera, and an anthropomorphic robot. Outcomes highlight an Average Euclidean Distance value of 1.8 mm on the best-processed item in the YCB-Video Dataset. Similarly, an Average Euclidean Distance value of 1.9 mm is achieved on the ISS Dataset. Eventually, the measurement uncertainty of the six components of the 3D poses is inspected. The latter assessment is performed on the ISS Dataset by relying on the ISO GUM [1].
2020
2020 IEEE International Workshop on Metrology for Industry 4.0 & IoT
978-1-7281-4892-2
File in questo prodotto:
File Dimensione Formato  
3DOPEDL Accuracy Evaluation of a Deep Learning Framework for 3D Object Pose Estimation.pdf

Accesso riservato

: Publisher’s version
Dimensione 4.84 MB
Formato Adobe PDF
4.84 MB Adobe PDF   Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11311/1143105
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 2
  • ???jsp.display-item.citation.isi??? 1
social impact