Reconstruction Bottlenecks in Object-Centric Generative Models

TL;DR

This paper investigates how reconstruction bottlenecks affect scene decomposition in object-centric generative models, revealing their critical role in reconstruction quality and model behavior, especially in complex real-world images.

Contribution

It provides an empirical analysis of reconstruction bottlenecks in GENESIS, highlighting their impact on object discovery and segmentation in complex datasets.

Findings

Reconstruction bottlenecks influence segmentation quality.

Bottlenecks determine the success of scene decomposition.

Model behavior is critically affected by the design of reconstruction constraints.

Abstract

A range of methods with suitable inductive biases exist to learn interpretable object-centric representations of images without supervision. However, these are largely restricted to visually simple images; robust object discovery in real-world sensory datasets remains elusive. To increase the understanding of such inductive biases, we empirically investigate the role of "reconstruction bottlenecks" for scene decomposition in GENESIS, a recent VAE-based model. We show such bottlenecks determine reconstruction and segmentation quality and critically influence model behaviour.