As per the literature research here:

https://docs.google.com/spreadsheets/d/1dyR4vvBzndGq99N2MnBAB63vkKE827IGJ48lGDgI8_k/edit#gid=0

The patterns that could be interesting to visualize are:

Blink duration
Blink duration 50_50
Amplitude (Max Height)
Lid closure speed
The peak closing velocity
The peak opening velocity
Delay of eyelid reopening
Duration at 80%
Closing time
Opening time
Eye gaze
PERCLOS
Average eye closure speed
Blink frequency
maximum close duration
average opening level
Eye closed duration

Currently, every time the analysis is performed, the videos needs to be analyzed using teyered (https://github.com/K-Kielak/teyered) framework from scratch to extract eyes closedness. This is the step that takes vast majority of the time spent on the analysis (for 2h session takes over 4.5h to be processed).

Once eye closedness is extracted, the rest of the analysis can be performed in a matter of seconds/minutes. Therefore, to highly speed up re-analysing the videos, there should be an option to save extracted eye closedness and re-use it instead of performing extraction from scratch each time the session is to be analyzed.

Currently all the videos from the RLD dataset have various formats and are very large (> 1gb) a compression with ffmpeg using something like:

ffmpeg -i $video_class -vcodec libx265 -crf 28 video/$person_id/$class_id'_compressed.mp4'

Which should reduce it to < 100 MB.

Fitting polynomial into data doesn't really represent the total data trend very well (they are limited in their expresiveness). The better, easier to interpret solution would be to plot moving average with a confidence area (one standard deviation up, one standard deviation down).

This will show exact, smoothed trend of the data and changes in it's variance - both things can give some tiredness signs and are not visible in current plots.

Specifically for: https://sites.google.com/view/utarldd/home

There exist some big outliers. Thus, matplotlib trying to show all data (including outliers) uses very wide spread between lower and upper limits of axes (Y in case of data, X in case of histograms). This causes that the actual useful body of data occupies very little of the plot making it less readable. Eliminating outliers from the data before plotting it using matplotlib is required.

Since videos have been processed in WP01 and WP02, it is time to extract their features together with appropriate labels and put them in one big CSV for easy of further analysis.

Add trends to existing plots and possibly add some histograms that explain data better.