VISAtlas: An Image-Based Exploration and Query System for Large Visualization Collections via Neural Image Embedding

Area of Science:

• Computer Science
• Information Visualization

Background:

• Existing visualization collection exploration systems rely on extrinsic attributes (e.g., authors, publication years).
• These systems neglect intrinsic visual properties, hindering effective visual comparison and design retrieval.
• There is a need for systems that leverage visual appearance for deeper insights into visualization collections.

Purpose of the Study:

• To present VISAtlas, an image-based approach for exploring and querying visualization collections using neural image embedding.
• To enable multi-perspective exploration and design retrieval based on the visual appearance of visualizations.
• To facilitate comparative analysis and discovery of novel visualization designs.

Main Methods:

• Developed a comprehensive dataset of synthetic and real-world visualizations.
• Trained a convolutional neural network (CNN) model with triplet loss for visualization classification.
• Designed a coordinated multiple view (CMV) system with a novel embedding overview, density plots, and sampling techniques.
• Utilized contextual layout frameworks to preserve embedding vector context and visualization taxonomies.

Main Results:

• VISAtlas effectively supports comparative analysis, exploration of composite visualizations, and image-based retrieval.
• Real-world visualization collections (e.g., Beagle, VIS30K) demonstrate richer diversity than synthetic ones (e.g., Data2Vis).
• Inspiring composite visualizations were identified within real-world collections, revealing distinct design patterns across sources.

Conclusions:

• VISAtlas enhances the exploration and understanding of visualization collections by incorporating visual appearance.
• Real-world data is crucial for capturing the true diversity and complexity of visualization designs.
• The system facilitates the discovery of new visualization designs and patterns through image-based retrieval.