Learning Visual Importance for Graphic Designs and Data Visualizations
summary
Models predict pixelwise importance from bitmaps using a VGG-16 fully convolutional network trained on crowdsourced BubbleView clicks (1400 MASSVIS images) and explicit annotations from the GDI dataset, running in ~100 ms per image and providing ~100x speedup over prior optimization methods. Predictions outperform natural-image saliency and the automatic variant of O'Donovan et al., enable retargeting and thumbnailing (user studies with hundreds of MTurk participants show 40% fewer clicks in a search task), and power an interactive realtime design tool. BubbleView clicks GDI dataset
AI-written summary · about
chapters (15)
- 0:00Introduction and motivation
- 1:06Visual importance definition
- 3:10Related work and datasets
- 4:15Measuring importance with eye tracking
- 6:05Crowdsourcing attention with BubbleView
- 7:52Explicit importance annotations
- 9:39Model architecture and training
- 11:02Evaluating visualization predictions
- 12:25Evaluating graphic design predictions
- 14:11Applications in retargeting
- 15:17Thumbnailing for visualizations
- 16:22Testing design variations
- 17:04Interactive design tool demo
- 17:46Future autonomous design
- 18:11Q&A session
related talks
38mWhere are we looking? Predicting human gaze using deep networks.
01h 18mComputer Vision after the Victory of Data
32mData Viz for Engineers: Optimizing Insight & Decision Making Through Visualization
01h 00mData, visualization, and designing AI
51mVisual Quality Assurance Using Vision-Language Models
50mDVUX: Data Visualization and User Experience
48mInterpretability - now what?
33mMachine Learning in Visual Search