GENESIS: Generative Scene Inference and Sampling with Object-Centric Latent Representations

TL;DR

GENESIS is a novel object-centric generative model for 3D scenes that captures object relationships, enabling scene decomposition and principled sampling of new scenes, advancing visual scene understanding in robotics and reinforcement learning.

Contribution

It introduces GENESIS, the first model to explicitly incorporate object relationships in 3D scene generation and decomposition, improving over prior models like MONet and IODINE.

Findings

Effective scene decomposition demonstrated on multiple datasets.

Capable of generating realistic novel scenes with object interactions.

Improved semi-supervised learning performance.

Abstract

Generative latent-variable models are emerging as promising tools in robotics and reinforcement learning. Yet, even though tasks in these domains typically involve distinct objects, most state-of-the-art generative models do not explicitly capture the compositional nature of visual scenes. Two recent exceptions, MONet and IODINE, decompose scenes into objects in an unsupervised fashion. Their underlying generative processes, however, do not account for component interactions. Hence, neither of them allows for principled sampling of novel scenes. Here we present GENESIS, the first object-centric generative model of 3D visual scenes capable of both decomposing and generating scenes by capturing relationships between scene components. GENESIS parameterises a spatial GMM over images which is decoded from a set of object-centric latent variables that are either inferred sequentially in an amortised fashion or sampled from an autoregressive prior. We train GENESIS on several publicly available datasets and evaluate its performance on scene generation, decomposition, and semi-supervised learning.