Performance tests

[inspirit]

Hi, i was wondering if you did any tests to measure how it performs against well known detectors,
such as OpenCV VJ etc. would be great to see some numbers
thanx

[luoyetx]

I would like to test the performance on FDDB and I will write the test code. Current training part code still need some modification, and it seems pretty hard to complete the training process over T*K=5400 CART, hard negative mining is consuming too many background images and need a solution to solve this problem, maybe more background images.

[inspirit]

just wondering how many negative images you are using? and what is the average image size (width x height)?

[luoyetx]

I modified the model with 5 landmarks ( 27 landmarks on paper). I also prepared 20000 background images (average size is 500 x 500 I guess) and 10000 face images. 5 landmarks is not good for feature extraction and the classifier is too weak to detect face in a image. But I didn't meet the problem which background images is insufficient. However, those using 27 landmarks have meet the problem!

[inspirit]

i think 5 landmarks still can do a good job and using less landmarks will produce more compact model file in the end. in my tests i see that LBF based alignment performs quite bad when the amount of landmark is small.
it clearly loose to ERT: http://www.cv-foundation.org/openaccess/content_cvpr_2014/papers/Kazemi_One_Millisecond_Face_2014_CVPR_paper.pdf

when you do alignment of 29+ landmarks they are quite close in precision.
I already posted about trying NPD features. i would also recommend to use ERT for landmarks alignment instead of LBF. it should improve performance with 5 landmarks

[luoyetx]

I will consider to use NPD + ERT, but I would like to try 27 landmarks first. 😃

Thanks a lot for your suggestion!

[ghost]

In “Face alignment using cascade gaussian process regression trees”CVPR2015 the author has told LBF is better than ert, so I think we should test first.