Skip to main navigation Skip to search Skip to main content

Scale Semantic Flow Preserving Across Image Pyramid

  • Zhili Lin
  • , Guanglu Song
  • , Biao Leng*
  • *Corresponding author for this work
  • Beihang University

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

Abstract

Image pyramid based face detector is powerful yet time consuming when cooperated with convolutional neural network, and thus hard to satisfy the computational requirement in real-world applications. In this paper, we proposed a novel method using Semantic Preserving Feature Pyramid (SPFP) to eliminate the computational gap between image pyramid and feature pyramid based detectors. Since a feature tensor of an image can be upsampled or downsampled by Reversible Scale Semantic Flow Preserving (RS2FP ) network, we do not need to feed images with all scales but a middle scale into the network. Extensive experiments demonstrate that the proposed algorithm can accelerate image pyramid by about 5 × to 7 × on widely used face detection benchmarks while maintaining the comparable performance.

Original languageEnglish
Title of host publicationNeural Information Processing - 28th International Conference, ICONIP 2021, Proceedings
EditorsTeddy Mantoro, Minho Lee, Media Anugerah Ayu, Kok Wai Wong, Achmad Nizar Hidayanto
PublisherSpringer Science and Business Media Deutschland GmbH
Pages464-471
Number of pages8
ISBN (Print)9783030923068
DOIs
StatePublished - 2021
Event28th International Conference on Neural Information Processing, ICONIP 2021 - Virtual, Online
Duration: 8 Dec 202112 Dec 2021

Publication series

NameCommunications in Computer and Information Science
Volume1516 CCIS
ISSN (Print)1865-0929
ISSN (Electronic)1865-0937

Conference

Conference28th International Conference on Neural Information Processing, ICONIP 2021
CityVirtual, Online
Period8/12/2112/12/21

Keywords

  • Face detection
  • Scale attention
  • Semantic preserving

Fingerprint

Dive into the research topics of 'Scale Semantic Flow Preserving Across Image Pyramid'. Together they form a unique fingerprint.

Cite this