Keywords:Deep Generative Models, World Models, Scene Interpretation
Ability to decompose complex environment which include many objects into individual component based on its semantic or functional structure is important ability in our higher-order cognition.Recently,researches about “World Models” that are models of surrounding environment to predict future states have gained much attention.This study aims at advancing such models considering object recognition.Prior works of scene interpretation using generative models are conducted under fully-unsupervised manner. However, this makes the problem ill-posed and the decomposition results do not always become as we intended.In this research, we incorporate knowledge about target into consideration, and develop a method that can decompose scenes include complex objects. Specifically, we develop a model that contrast distributions of foreground and background to enable arbitrary decomposition, and we show that this method is capable of decompose challenging datasets that previous methods cannot.
Authentication for paper PDF access
A password is required to view paper PDFs. If you are a registered participant, please log on the site from Participant Log In.
You could view the PDF with entering the PDF viewing password bellow.