Marks, Channels, and Dead Ends: Stop Running Graphical Perception Studies and Start Modeling Visualizations as Images

A new line of attack on visualization evaluation is making the rounds in HCI and graphics circles: stop treating chart readability as an encoder problem and start treating it as an image problem. The paper argues that graphical perception studies, long used to rank marks and channels, can miss how people actually perceive a visualization once the full rendered image hits the visual system.

The practical complaint is familiar to anyone who has tuned a chart for production: tiny shifts in data distribution, point density, axis scaling, or other micro-design choices can change what users perceive even when the nominal encoding stays the same. That means a guideline derived from one controlled study may not survive contact with a different dataset or a slightly different layout.

The proposed alternative is to evaluate visualizations with computational models of human vision that take pixels as input. In that framing, the important unit is the retinal image and the emergent structure it creates, not just the abstract data-to-mark mapping. The authors say this approach is more scalable, more grounded in perception, and better at catching effects that traditional encoding theory misses.

For developers building dashboards, analytics tools, or any data-heavy UI, the takeaway is straightforward: don’t assume a “best practice” encoding will stay best once the real data and rendering pipeline get involved. The paper also claims image-based vision models can reproduce established scatterplot results while predicting discriminability, which makes this a potentially useful bridge...