Data set representativeness and skewed accelerations
by dymitrruta - Thursday, April 2, 2015, 08:53:15
Dear Organisers, I have 3 questions to be clarified please if possible: First, I noticed significant deviations between the training and testing sets, which I guess is explained by the different firefighters data in both sets. Could you please confirm that invariably this will be the case for the final test i.e. none of the firefighters involved in the training set will be present in the final testing set? Second, I wonder if it is possible to know whether all the classes and their combinations observed in the training set are equally represented in the testing set, i.e. if the prior class probabilities and their combinations are similar for both sets? Third, while most of the accelerations observed are oscylatory around 0, I noticed in many cases acceleration time series showing large stationary numbers say -9m/s2. Is it possible that some accelerometers were faultily shifted i.e not scaled properly? It is unlikely to expect the body parts moving by 10s of metres in a span of a second or so in a perfectly free-fall like move?
Re: Data set representativeness and skewed accelerations
by mich - Thursday, April 02, 2015, 16:44:19

Dear Dymitr, 

ad 1.: YES

ad 2.: YES

ad 3.: we do not compensate gravitational force from raw readings, so -9m/s2 sound more or less right for non-movement (z-axis pointed downard). Please note that fast movemt of feet for example can easily achive 160 m/s2.
As for oscilation around zero... we were looking into the data very closely before opening them into competiotion and I woudn't say that acceleration oscilates around 0 (maybe you are looking into gyroscopes?)