Mixture of Cluster-Guided Experts for Retrieval-Augmented Label Placement

Area of Science:

• Computer Vision
• Graphic Design
• Machine Learning

Background:

• Text labels are crucial for conveying information in visualizations and graphic design.
• Existing label placement methods struggle with diverse label layouts and fail to leverage design knowledge from prior work.
• Variability in labeled object categories and structures creates complex layout challenges.

Purpose of the Study:

• To develop an advanced label placement solution that addresses the limitations of current single-model approaches.
• To effectively capture fine-grained differences in diverse label layouts.
• To integrate and exploit embedded design knowledge from existing label layouts.

Main Methods:

• Proposed a Mixture of Cluster-guided Experts (MoCE) approach for label placement.
• Implemented the Label Placement Cluster-guided Experts (LPCE) model, featuring a MoCE layer with multiple feed-forward networks (FFNs) as experts.
• Introduced a retrieval augmentation strategy to enrich sample representations with reference layouts.

Main Results:

• The LPCE model demonstrated superior quantitative and qualitative performance in label placement tasks.
• LPCE surpassed various state-of-the-art baseline methods.
• The cluster-based gating function effectively assigns samples to specialized experts.

Conclusions:

• The proposed LPCE model offers a significant advancement in automated label placement.
• Leveraging cluster-guided experts and retrieval augmentation enhances the model's ability to handle complex layouts.
• The approach effectively integrates design knowledge, leading to improved placement outcomes.