Bridging Depth Estimation and Completion for Mobile Robots Reliable 3D Perception

Dimitrios Arapis*, Milad Jami, Lazaros Nalpantidis

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingArticle in proceedingsResearchpeer-review

Abstract

Autonomous mobile robots rely on knowing the 3D structure of their environment in order to plan and operate safely. This paper addresses the problem of estimating depth maps from monocular RGB images and any potentially available sparse depth measurements—as would be the case of using inexpensive RGBD or LIDAR sensors. Our approach bridges depth estimation and depth completion by introducing a novel lightweight two-stream encoder-decoder neural network that exploits late fusion and additional output refinements. The novelty of our network is that it can benefit from any available depth data, while not making any assumptions about their actual availability, density or distribution. Our proposed architecture outperforms models with equivalent number of operations in terms of depth accuracy and performs on par with much larger models.

Original languageEnglish
Title of host publicationRobot Intelligence Technology and Applications 7 : Results from the 10th International Conference on Robot Intelligence Technology and Applications
EditorsJun Jo, Han-Lim Choi, Marde Helbig, Hyondong Oh, Jemin Hwangbo, Chang-Hun Lee, Bela Stantic
PublisherSpringer
Publication date2023
Pages169-179
ISBN (Print)9783031268885
DOIs
Publication statusPublished - 2023
Event10th International Conference on Robot Intelligence Technology and Applications - Gold Coast, Australia
Duration: 7 Dec 20229 Dec 2022

Conference

Conference10th International Conference on Robot Intelligence Technology and Applications
Country/TerritoryAustralia
CityGold Coast
Period07/12/202209/12/2022
SeriesLecture Notes in Networks and Systems
Volume642 LNNS
ISSN2367-3370

Keywords

  • 3D perception
  • Depth completion
  • Depth estimation

Fingerprint

Dive into the research topics of 'Bridging Depth Estimation and Completion for Mobile Robots Reliable 3D Perception'. Together they form a unique fingerprint.

Cite this