Synthesize-It-Classifier: learning a generative classifier through recurrent self-analysis

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

1 Citation (Scopus)

Abstract

We show the generative capability of an image classifier network by synthesizing high-resolution, photo-realistic, and diverse images at scale. The overall methodology, called Synthesize-It-Classifier (STIC), does not require an explicit generator network to estimate the density of the data distribution and sample images from that, but instead uses the classifier's knowledge of the boundary to perform gradient ascent w.r.t. class logits and then synthesizes images using the Gram Matrix Metropolis Adjusted Langevin Algorithm (GRMALA) by drawing on a blank canvas. During training, the classifier iteratively uses these synthesized images as fake samples and re-estimates the class boundary in a recurrent fashion to improve both the classification accuracy and quality of synthetic images. The STIC shows that mixing of the hard fake samples (i.e. those synthesized by the one-hot class conditioning), and the soft fake samples (which are synthesized as a convex combination of classes, i.e. a mixup of classes [36]) improves class interpolation. We demonstrate an Attentive-STIC network that shows iterative drawing of synthesized images on the ImageNet dataset that has thousands of classes. In addition, we introduce the synthesis using a class conditional score classifier (Score-STIC) instead of a normal image classifier and show improved results on several real world datasets, i.e. ImageNet, LSUN and CIFAR 10.

Original languageEnglish
Title of host publicationProceedings - 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2021
EditorsMargaux Masson-Forsythe, Eric Mortensen
Place of PublicationPiscataway NJ USA
PublisherIEEE, Institute of Electrical and Electronics Engineers
Pages5157-5166
Number of pages10
ISBN (Electronic)9781665445092
ISBN (Print)9781665445108
DOIs
Publication statusPublished - 2021
EventIEEE Conference on Computer Vision and Pattern Recognition 2021 - Online, Virtual, Online, United States of America
Duration: 19 Jun 202125 Jun 2021
https://cvpr2021.thecvf.com/ (Website)
https://ieeexplore.ieee.org/xpl/conhome/9577055/proceeding (Proceedings)

Publication series

NameProceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition
PublisherIEEE, Institute of Electrical and Electronics Engineers
ISSN (Print)1063-6919
ISSN (Electronic)2575-7075

Conference

ConferenceIEEE Conference on Computer Vision and Pattern Recognition 2021
Abbreviated titleCVPR 2021
Country/TerritoryUnited States of America
CityVirtual, Online
Period19/06/2125/06/21
Internet address

Cite this