Skip to content

Latest commit

 

History

History
 
 

Metrics

Metrics

Three different metrics are computed, with a focus on efficiency and human scanpath similarity and prediction:

  • Cumulative performance (AUCperf): Proportion of targets found (vertical axis) for a given number of fixations (horizontal axis). The Area Under the Curve is computed. To measure the similarity with humans, the final score is computed by: 1 - |AUCperf(subjects) - AUCperf(model)|.
  • MultiMatch (AvgMM and Corr): Compares two given scanpaths in several dimensions, by treating them as geometrical vectors in a two-dimensional space. Models' scanpaths are compared against those of human subjects and the average is computed across all dimensions, with the exception of time. The correlation between the Multi-Match scores of models against participants (hmMM) and the Multi-Match scores of subjects against other subjects (whMM) is also reported. See https://multimatch.readthedocs.io for more information on the algorithm.
  • Human Scanpath Prediction (AUChsp, NSShsp, IGhsp and LLhsp): Given the scanpath of a human subject, each model attempts to predict where the next fixation is going to land. This is done for each fixation in the scanpath (with the exception of the first one) and allows for the computation of the Area Under the Curve (AUC), Normalized Scanpath Saliency (NSS) and Information Gain relative to the center bias model (IG) and the uniform model (LL). Results are averaged across all fixations. This metric originated from the pre-print State-of-the-art in Human Scanpath Prediction.