Unsupervised co-generation of foreground-background segmentation from Text-to-Image synthesis

Yeruru Asrar Ahmed, Anurag Mittal

Abstract: Highlights•We propose a novel GAN framework incorporating a co-segmentation approach.•ForeGround-BackGround (FG–BG) mask prediction and synthetic image generations.•Our proposed approach achieves superior mask generation quality.•Generate masks in unconditional settings in image synthesis.•Dual branch prediction is incorporated to improve mask generation quality.•Demonstrate performance enhancement achieved usin Text-to-Image synthesis for mask generations.