Performance on Columbia Gaze Dataset?

[saturosfz]

Anyone who run the experiment? My result is terrible....

[TadasBaltrusaitis]

For Columbia I resized the images to 800 x 533 pixels for Columbia as otherwise it was taking too long to run and was getting false positives for face detection, this helped somewhat with gaze estimation.

However, the results on Columbia were not as accurate as you mentioned (on a subset I get ~12 degrees mean and ~11 degrees median absolute error).

At the moment gaze detection is the weakest part of OpenFace, but I am working on some improvements.

Thanks,
Tadas

[rishubjain]

Hi,

Will these improvements to gaze detection be released soon? (and how soon?)

Thanks!

[TadasBaltrusaitis]

This is a much longer term project and the timeline is more within 6 months to a year.

[yyxiongzju]

Hi @TadasBaltrusaitis,

How did you get the mean error and absolute error gaze estimation for Columbia dataset?
It seems no ground truth for gaze direction in Columbia dataset which is the way Openface estimates.
Thanks.

[TadasBaltrusaitis]

You can compute the gaze direction from the Columbia dataset by either using the gaze angles or actually computing the eye gaze vector from the location of the eye to the target they are looking at.

[cunesewangst]

@TadasBaltrusaitis
hello TadasBaltrusaitis, Do you have the code for pre-processing columbia gaze dataset? thank you