Investigating object compositionality in Generative Adversarial Networks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep generative models aim to replicate data generation processes for synthesis and representation extraction.
• Existing image generation models often overlook the compositional structure inherent in human visual perception.
• Object compositionality, a key aspect of human scene understanding, has been an underutilized inductive bias in generative models.

Purpose of the Study:

• To investigate object compositionality as an inductive bias for Generative Adversarial Networks (GANs).
• To develop a modified GAN architecture that explicitly incorporates object compositionality.
• To evaluate the effectiveness of structured GANs in generating realistic multi-object images and learning meaningful representations.

Main Methods:

• A minimal modification to standard GAN generators was proposed to integrate object compositionality.
• Two extensions were developed to model dependencies between objects and backgrounds.
• The approach was evaluated on multiple multi-object image datasets, including the CLEVR dataset.

Main Results:

• The modified GANs reliably learned to generate images as compositions of objects.
• Structured GANs demonstrated improved fidelity in generating multi-object images compared to the reference distribution.
• The learned generative process structure enabled unsupervised instance segmentation by inverting the model.
• The proposed approach outperformed recent unsupervised object-centric methods on the CLEVR dataset.

Conclusions:

• Incorporating object compositionality as an inductive bias significantly enhances GANs for multi-object image generation.
• Structured GANs offer a powerful framework for representation learning and enable novel applications like unsupervised instance segmentation.
• This work highlights the importance of human-like structural biases in advancing deep generative models.