Toward Improving the Generation Quality of Autoregressive Slot VAEs

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Unconditional scene generation from object-centric representations (slots) is challenging.
• Existing slot-based models struggle with learning multiobject relations for coherent scene imagination.

Purpose of the Study:

• To improve unconditional scene generation by strengthening object correlation learning.
• To develop methods for more effective joint learning of scene inference and generation.

Main Methods:

• Conditioning slots on a global, scene-level variable to capture higher-order correlations.
• Learning a consistent object order for autoregressive generation of scene objects.
• Training an autoregressive slot prior for sequential object generation.

Main Results:

• Demonstrated clear gains in unconditional scene generation quality across three multiobject environments.
• Ablation studies validated the effectiveness of the proposed improvements in object correlation learning.

Conclusions:

• The proposed methods significantly enhance the ability of slot-based models to learn object correlations.
• Improved object correlation learning leads to higher quality unconditional scene generation.