Relativistic Discriminator Based Unsupervised Pose Guided Human Image Generation Using Gan

Abstract

Generative Modeling Is Nowadays Becoming Popular As Generative Adversarial Networks And Variational Autoencoders Are Producing Great Results In Image Generation, One Of Which Is To Generate Images Of The Human Body From Poses (3D Sketches). However, Most Of The Existing Works Are Requiring To Use Paired Set Of Datasets Which Are In Some Cases Very Hard To Obtain And Complex. To Overcome This Problem, One Approach Is To Use An Unsupervised Mechanism For Unpaired Data Sets. This Research Work Aims At Introducing Generative Modeling That Generates A Human Image That Gives Control On The Type Of Clothing And Poses Provided By The User In An End To-End Unsupervised Manner Using Cyclegan And Variational Autoencoder. Further, The Previous Works Suffer In Capturing The Details Of Clothing Colors And The Overall Generated Image Suffers From Artifacts And Getting The Desired Pose. So, To Alleviate These Problems This Research Work Introduces Color Cycle Consistency Loss, Complementary Loss, And Relativistic Discriminator To The Cyclegan Model To Generate Visually More Realistic Images. Moreover, The Inception Score Of The Proposed Solution Along With Different Experiments Have Been Measured And The Proposed Solution Has Shown Improvements In Generating The Desired Image With Better Quality And Pose Informationwith The Score Of 3.55 And 3.01 Respectivelly On The Upper And Full Body . Further, Human Evaluation Was Held To Evaluate The Quality Of The Images Being Generated And It Has Shown The Proposed Solution Generates Better Quality Images With Average Of 75% And 66.64% Respectively For The Upper And Full Body. Finally, From The Evaluation Results, The Added Relativistic Discriminator, Color Cycle Consistency Loss, And Complementary Loss Have Improved The Quality Of The Image Being Produced.