Visualization: Reading Detection in Real-Time

Global Prediction

Local Prediction

This user task was: Reading

Our RRSVM Prediction was: Reading

About This Data

The visualizations above are fully labeled examples from our test set.

We use our RRSVM framework to optimize both globally (Left) and locally (Right) and provide the resulting labeled fixations above.

Here blue represents a fixation point in which the user was predicted to be reading, and red when they were predicted to be skimming.

The overall document level prediction, given by the globally optimized RRSVM is provided below the predictions.

To produce the data, users were shown various articles from cnn.com/health and tasked to either read or to skim. These global labels were propagated to label the local windows.

For more information see our ETRA '19 paper in the publications section of this Web page, or see the forward below:


Reading detectors operate by inputting windows of sequential eye fixations and outputting predictions of the fixation behavior during those windows as being reading or skimming.

Here we introduce a new method for reading detection using the Region Ranking SVM (RRSVM). The high level idea is that the RRSVM combines an SVM clasisfier on the local features of the fixation windows with global knowledge gained from the fixation samples most representative of reading and skimming across stimuli. The RRSVM is able to produce global , document-level and local window-level labels. The detector is able to offer both with only coarse grained global ground truth as opposed to labrious labeling of individual fixation windows obtained through manual inspection. We offer a flexible framework to optimize for the former or the later. As the detector provides fixation level labels, it can be used for real-time applciations that require short prediction delays.

The RRSVM reading detector accurately predicted 82.5% of the global (article-level) reading/skimming behavior, with accuracy in predicting local window labels ranging from 72-95%, de- pending on how tuned the RRSVM was for local and global weights.