the result in LFPW and Helen

[yyytq]

first of all thank you for your code ! it works well ! but i notice that when i run it on LFPW, the err is much higher than the paper(5.17/3.35). but on Helen, the result is nearly the same . how do you think about this issue? the second one i'm confusion is that for helen (194 landmakers), the paper use total 300 trees, averaged less than 2 for each maker. how it comes??? it seems like only decision tree instead random forest~~~~(>_<)~~~~

looking forward to your reply~~~ thanks again!

[jwyang]

@yyytq ,

but i notice that when i run it on LFPW, the err is much higher than the paper(5.17/3.35). but on Helen, the result is nearly the same . how do you think about this issue?

Hi, on LFPW dataset, do you test also using 29 landmarks as used in the paper? If not, the error should be 5.00+

the second one i'm confusion is that for helen (194 landmakers), the paper use total 300 trees, averaged less than 2 for each maker. how it comes??? it seems like only decision tree instead random forest~~

Yes, I am also a little confused on that. it is indeed a little odd just using 300 trees for 194 landmarks, i think there are some tricks on that, it would be better to ask the authors, :)

[yyytq]

thank you very much!! I have tested 29 landmarkrs . what i'm thinking about is that on LFPW the result is 5.00+/3.35, on helen is 5.5+/5.81,5.4 why the code perform much better on helen? is only the reason of the tree number?

what's more, do you think the different raio_radius may influence a lot ?

[jwyang]

Oh, I see, so what is the training sample you used for testing each dataset. the tree number for these two sets are same to each other, right? In this case, I think the size of training sample made the difference, possibly

Not that much, I set them empirically actually

On Thursday, November 26, 2015, yyytq [email protected] wrote:

thank you very much!! I have tested 29 landmarkrs . what i'm thinking about is that on LFPW the result is 5.00+/3.35, on helen is 5.5+/5.81,5.4 why the code perform much better on helen? is only the reason of the tree number?

what's more, do you think the different raio_radius may influence a lot ?

— Reply to this email directly or view it on GitHub https://github.com/jwyang/face-alignment/issues/26#issuecomment-159825632 .

[yyytq]

i just use the trained model you have provided ! i means the paper use less trees on helen, but we use the same trees number as LFPW . so if our result on helen is really compareable with the paper ?

[jwyang]

yes, you are right, that is one of the difference. Another difference is that the model has trees with depth=4, while the paper has trees with depth=5, maybe you can modify the parameters to make them similar to the paper and have a try again.

[yyytq]

yeah, i have tried that. in fact, what i want to do now is change the initial shape from meanshape to a shape that have been per-processed, so that the err can be reduced. but i always can't get good result (>_<) do you think it is meaningful?

[jwyang]

it is a difficult thing, because at the begining, we do not know exactly the pose and expression of a face. Actually, the first stage can be regarded as a pre-processing step for the following face alignment. I also did some experiment to estimate a more plausible initial shape before but failed to imporve the performance. As a result, I do not think changing another initial shape does not make much different to the final alignment result.

On Fri, Nov 27, 2015 at 2:41 AM, yyytq [email protected] wrote:

yeah, i have tried that. in fact, what i want to do now is change the initial shape from meanshape to a shape that have been per-processed, so that the err can be reduced. but i always can't get good result (>_<) do you think it is meaningful?

— Reply to this email directly or view it on GitHub https://github.com/jwyang/face-alignment/issues/26#issuecomment-160060336 .