Synthetic to Real Gap Estimation of Autonomous Driving Datasets using Feature Embedding

Nivesh Gadipudi, Irraivan Elamvazuthi*, Mahindra Sanmugam, Lila Iznita Izhar, Tindyo Prasetyo, R. Jegadeeshwaran, Syed Saad Azhar Ali

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Recent advances in autonomous driving using deep learning have drawn immense attention from robotics and computer vision communities. Training generalized deep learning models for autonomous driving tasks like visual odometry, segmentation, and object detection requires large amounts of data. Acquiring real-world data with accurate annotations is time-consuming and expensive. Due to this challenge, synthetic datasets are increasingly being used for training and testing deep learning models. Synthetic data lacks the appearance and contextual properties of real-world datasets. Several works have been shown to reduce this gap between synthetic and real-world images. However, evaluating the gap between the synthetic and real-world datasets is a longstanding challenge because of its highly not deterministic nature. This research proposes the use of feature embedding techniques to address this synthetic to reality gap in the form of distance between different data clusters. From the experiments, the proposed approach estimated the distance between real-world to enhanced virtual datasets is 6-10 times the distance between real-world to virtual datasets.

Original languageEnglish
Title of host publication2022 IEEE 5th International Symposium in Robotics and Manufacturing Automation, ROMA 2022
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9781665459327
DOIs
StatePublished - 2022
Event5th IEEE International Symposium in Robotics and Manufacturing Automation, ROMA 2022 - Malacca, Malaysia
Duration: 5 Aug 02027 Aug 0202

Publication series

Name2022 IEEE 5th International Symposium in Robotics and Manufacturing Automation, ROMA 2022

Conference

Conference5th IEEE International Symposium in Robotics and Manufacturing Automation, ROMA 2022
Country/TerritoryMalaysia
CityMalacca
Period5/08/027/08/02

Bibliographical note

Publisher Copyright:
© 2022 IEEE.

Keywords

  • Autonomous driving
  • feature embedding
  • reality gap
  • synthetic datasets

ASJC Scopus subject areas

  • Artificial Intelligence
  • Industrial and Manufacturing Engineering
  • Control and Optimization

Fingerprint

Dive into the research topics of 'Synthetic to Real Gap Estimation of Autonomous Driving Datasets using Feature Embedding'. Together they form a unique fingerprint.

Cite this