pupil
pupil copied to clipboard
fixation detector better default values
A user reports the at the current default miss a lot of fixations. Even with high confidence data we see second long stretches without any fixations.
Please investigate.
The high false negative rate is most likely caused by the assumption that there are no outliers. To improve the algorithm, I will use the mean of the 20% most biggest pair-wise cosine distances as dispersion measure. This way we reject the fixation if there are few extreme outliers or if a lot of the vectors exceeding the maximum dispersion.
I will also increase the maximum dispersion default to 3 degrees. This has been the default value for the online detector as well.