evgeniimakarov

Regarding differences in performances of facial landmark detector combined with different face detectors you can check this paper https://arxiv.org/abs/1511.05049 In few words, dlib landmark detector isn't very robust to random bbox shifts although very robust to scale of bbox. That means you can pick any face detector you like unless it deviates in x-y plane too much from dlib (pls note dlib face detector trims chin of person)

Better keep them. We do that and nothing bad happens. And shape regressor learns people with glasses

Accordingly to recent VOT2017 tracker competition there was huge progress in recent years in tracker design. For example winner of VOT2014 DSST algorithm now takes 47th place out of 51 candidate (http://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w28/Kristan_The_Visual_Object_ICCV_2017_paper.pdf). At the same time DSST is tracking algorithm currently implemented in dlib. Is there any possibility to introduce new tracker in dlib? CSRDCF++ seems to be good option since it showed best results...

Perhaps you'd better look into EOS library from Patrick Huber to find orientation of face from 2D landmarks

I want to train my own face detector using "fhog_object_detector_ex.cpp" and use it for precise face detection with smartphone camera in real-time. I red examples and ran through discussions here and on github but still can't get some nuances and want to get some valuable feedback. My dataset is about 20K selfie pictures collected from real users in hard light conditions and with pose variations. They are 224x224 with faces occupying area of about 80x80 - 180x180 pixels. What is good with them is...

I want to train my own face detector using "fhog_object_detector_ex.cpp" and use it for precise face detection with smartphone camera in real-time. I red examples and ran through discussions here and on github but still can't get some nuances and want to get some valuable feedback. My dataset is about 20K selfie pictures collected from real users in hard light conditions and with pose variations. They are 224x224 with faces occupying area of about 80x80 - 180x180 pixels. What is good with them is...

Finally I was able to get it. Works properly. Thanks a lot!

Yes, it works nicely although a bit slowly. objective: 1.86599 objective gap: 0.00844593 risk: 0.0192628 risk gap: 0.00844593 num planes: 102 iter: 206 BTW, my dataset is 20 000 images. 206 iterations on 8 images may turn to billions for my dataset :) may this serve as RAM issue?