search

chili-epfl / attention-tracker

136 stars 96 forks source link

node to compute online with_me_ness with cowriter #2

Closed alexis-jacq closed 8 years ago

alexis-jacq commented 8 years ago

We measure with-me-ness by comparing the attentional focus of the human (as estimated in real-time by the robot) with the expected, a priori targets of attention elicited by the task at hand.

When the target of the human's focus is one of the expected, the with-me-ness is incremented. Other with it is decremented. If for some reason the face of the human is not detected (so we have no information about the target, the value of with-me-ness is conserved.

In this algorithm we compute this value using an exponential moving average approach: with-me-ness is incremented/decremented as follow (respectively) :

with-me-ness = (1-mu)_with-me-ness + mu with-me-ness = (1-mu)_with-me-ness

The more mu is close to 0, the more the robot memorize the last events and the lower the value will move. If with-me-ness is initialized at 0.5, this value stays between 0 and 1.

severin-lemaignan commented 8 years ago

thanks for the PR! Very useful!

Regarding the computation of the with-me-ness, why not relying on the algorithm proposed in the paper submitted to HRI? https://github.com/chili-epfl/paper-HRI2016-attentionTracker If the paper is accepted, this would allow to have a 'scientifically validated' measure, no?

For reference, here is the code I have used to compute the with-me-ness values reported in the paper.

alexis-jacq commented 8 years ago

This is directly inspired from the algorithm in https://github.com/chili-epfl/paper-HRI2016-attentionTracker. I just chose to use a moving average instead of an average on the whole session. In the paper you defined a interval and computed the value on the interval. So this is the same approach, but with a moving interval..

I can also implement exactly the same algorithm (with two accumulators). The only thing : I think we should'nt take into account the instants where we loose a face : in these instants, we dont know if the child is with the robot or not.

2015-10-13 15:15 GMT+02:00 Séverin Lemaignan notifications@github.com:

thanks for the PR! Very useful!

Regarding the computation of the with-me-ness, why not relying on the algorithm proposed in the paper submitted to HRI? https://github.com/chili-epfl/paper-HRI2016-attentionTracker If the paper is accepted, this would allow to have a 'scientifically validated' measure, no?

For reference, here https://github.com/chili-epfl/paper-HRI2016-attentionTracker/blob/master/data/process.py#L156 is the code I have used to compute the with-me-ness values reported in the paper.

— Reply to this email directly or view it on GitHub https://github.com/chili-epfl/attention-tracker/pull/2#issuecomment-147710724 .

Alexis David JACQ, Doctoral Assistant Robotics Group, CHILI laboratory, EPFL