MPQ-YOLACT: Mixed-Precision Quantization for Lightweight YOLACT

Seungjin Lee, Hyun Kim

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

3 Scopus citations

Abstract

Segmentation network, which are widely used in various multimedia applications in recent years, are capable of precise image processing in units of pixels, but have higher computational complexity compared to other computer vision algorithms such as classification and object detection. In this paper, we propose a technique that applies mixed-precision quantization to the existing YOLACT network, which performs accurate instance segmentation. By adaptively applying quantization according to the parameter size and the effect on the accuracy of modules in YOLACT, it is possible to significantly reduce the network size while maintaining the accuracy of YOLACT as much as possible. The experimental results show the parameter size of the entire network is reduced by 75.4% with only a negligible drop in accuracy of less than 0.1%.

Original languageEnglish
Title of host publication2020 IEEE International Conference on Consumer Electronics - Asia, ICCE-Asia 2020
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9781728161648
DOIs
StatePublished - 1 Nov 2020
Event2020 IEEE International Conference on Consumer Electronics - Asia, ICCE-Asia 2020 - Seoul, Korea, Republic of
Duration: 1 Nov 20203 Nov 2020

Publication series

Name2020 IEEE International Conference on Consumer Electronics - Asia, ICCE-Asia 2020

Conference

Conference2020 IEEE International Conference on Consumer Electronics - Asia, ICCE-Asia 2020
Country/TerritoryKorea, Republic of
CitySeoul
Period1/11/203/11/20

Keywords

  • Deep learning
  • Instance segmentation
  • Mixed precision
  • Quantization
  • YOLACT

Fingerprint

Dive into the research topics of 'MPQ-YOLACT: Mixed-Precision Quantization for Lightweight YOLACT'. Together they form a unique fingerprint.

Cite this