Motion Cues and Everday Events
project update (better late than never?)
Its been a few months, but I've officially finished this project. The semester ended, and a job moved me away from the lab. I am working on the final write-up and will post it in the next week. All in all, the project was a success, and this blog was very useful for sharing information with my mentor and for keeping it organized for me.
analysis update
need new confidence levels with new duplos correlations and variance stuff
discussion ideas
what about an analysis of where the motion trackers were on the actual screen, related to segmentation. Question of media and media representation.
changes to Duphouse cuts; noise in videogame
In the midst of serious data analysis.. Found the following correction:
Duphouse cutpoints:
in 0:45:00, out 6:55:14.
Also, I found some weirdness in the videogame movie, which is not
catastrophic but adds some noise:
Looking at XYZ position with the movies. Clothes and duphouse look
reasonable. It looks like there might be a spot in videogame where
the right hand moved below the sensor and the signs flipped on
everything: between 13080 and 13100. Also right around 13050.
Finally, it looks like between 13150 and 13200 there was noise from
the TV.
R^2
raw number of units + 1. t-test (since we only have one variable).
R^2 -> squared correlation between what we predicted from the model (with linear combination), and the actual correct answer
correlation always positive.
squared correlation can be compared to amount of variability in depedent measure
data analysis overview
data analysis:
create a table with number of people clicking as a function of time, movie, and grain.
created graph of number of clicks per bin, based on amplitude could see that clicks tended to occur in same place for coarse. seemed true, to a lesser degree, for fine.
did a PCA analysis, and found that most of the variance was accounted for in the first eight components.
multiple linear regression;
dependent variable: number of people that tapped at fine grain. independent variables: 15 motion parameters.
worried about lag between event boundary and observation. Found greatest correlation between characteristic and event boundary for each event parameter, in seconds. results ranged from 1 s to 20 s. then realized negative lags could also exist. incorporated that.
forward stepwise regression. utilized this analysis to see which motion parameters sucked up the most variance. R^2
videogame motion characteristics, in order of importance for fine grain: LefSpeed, LeftHeadAccel, rightLeftSpeed, RighLeftDist, RightAccel
-> different motion parameters found for coarse grain
Hierarchical analysis
continuous distance. found minimum distance between a coarse unit and a fine unit.
expected distance from units by chance, based on formula in JP:general Found min 163.0 ms observed, compared to 443.0 ms expected. mean was about 3 seconds difference. substantial hierarchical effects.
bin analysis:
looked at whether bins of units overlap.
INTERESTING QUESTION: are there more fine coarse units for the negative time lagged correlations (e.g. in clothes)
Another interesting question: show participants the animation generated from motion data and ask them tosegment that. Problem: doesn't resemble action sequence too well.
graphs:
black = headspeed
red = segmentation activity
Movement regression and predictive features, with cumulative R^2 values:
Individual R^2 values (found by stepwise regression)
Note: we might need to do signifagance tests on the correlations between movements, because there is a large range (from -1 to 1).
Duphouse: left accel -> Head Accel
Videogame: Right speed -> head speed & accel
Clothes: right speed -> RightHead speed
Remaining results
-ANOVA analysis on number of breakpoints during eachv iewing, with interpreation condition as a between-participants variable and grain as a repeated measure.
- do two groups agree better about location of fine-grained boundaries than coarse-grained boundaries [breakpoint histograms are highly correlated for fine segmentation and less so for coarse segmentation]
Events Occurring After Duration of Movie
Some discrepancies I've found:
SubjectNumber: e4101
Condition: F
Movie: ListA
4 Item2 Item2 Item1 Item1 Item3 Item3 Item2 Practice 361468 1 N/A
This is off by 1 ms. At worst, I think we might be off by 1ms in our calculation of duration (some rounding must be used to go from frames to ms).
SubjectNumber: e4103
Condition: F
Movie: ListB
2 Item1 Item1 Item2 Item3 Item1 Item2 Item3 Practice 365220 1 N/A
This one is more serious. The events should not exist past 361467 and we're off by almost 4 seconds.
Duration and Sequences
It would be helpful to know the duration of each clip in (ms) to troubleshoot why there are 'events' after the end of clip. Also, knowing clip sequences:
Duration
N64: 240300 ms
Clothes: 498267 ms
Duplos: 361467 ms
Sequences
ListA::
IsList: True
Movie: "clothes.mov @ :Movies:" "duplos.mov @ :Movies:" "N64.mov @ :Movies:"
Levels: Item1 Item2 Item3
ListB::
IsList: True
Movie: "duplos.mov @ :Movies:" "clothes.mov @ :Movies:" "N64.mov @ :Movies:"
Levels: Item2 Item1 Item3
ListC::
IsList: True
Movie: "N64.mov @ :Movies:" "clothes.mov @ :Movies:" "duplos.mov @ :Movies:"
Levels: Item3 Item1 Item2
ListD::
IsList: True
Movie: "clothes.mov @ :Movies:" "N64.mov @ :Movies:" "duplos.mov @ :Movies:"
Levels: Item1 Item3 Item2
ListE::
IsList: True
Movie: "duplos.mov @ :Movies:" "N64.mov @ :Movies:" "clothes.mov @ :Movies:"
Levels: Item2 Item3 Item1
ListF::
IsList: True
Movie: "N64.mov @ :Movies:" "duplos.mov @ :Movies:" "clothes.mov @ :Movies:"
Levels: Item3 Item2 Item1
Trying to figure out why the are button clicks after the duration of a movie. The movies do indeed self-terminate in the Psyscope script.
Also, need to get the parameters used for Ascension data system that are stored on Abram's lab computer.
Quicktime movies have 29.97 fps
For N64,
Movie starts at 00:40:17.
00:15:21 corresponds to 12972.687 seconds, so we need to add:
00:40:17 - 00:15:21 = 24.8665332
Offset = 12972.687 + 24.867 = 12997.55353
For Clothes,
Movie starts at 00:42:06
00:08:25 --> 7336.669
00:42:06 - 00:08:25 = 33.3660327
Offset = 7336.669 + 33.3660327 = 7370.035033
For Duplos,
Movie starts at 00:54:00
00:08:25 --> 4437.839
00:54:00 - 00:08:25 = 45.13793103
Offset = (prev ans) + 4437.839 = 4482.976931
