Question about Detect result

[guo253]

Hi,
I'm so excited for your detect part.
when I runned according to your example,there is only Detect time showing(about 90 s),but nothing else.What's the problem? Isn't it should generate a alignned face in the face detected bbox?
Thanks again.

[denghuiru]

Hi， I have encountered the same problem with you, have you solved this problem yet ?

[guo253]

Just wait samuel -_-
@denghuiru

[denghuiru]

Hi, I have solved this problem, you can install imagemagick on your unbuntu, just use "apt-get install imagemagick", then you will see the result.

[denghuiru]

@guo253 use pt-get install imagemagick, run your code again, then you can see the result

[guo253]

@denghuiru I have install the imagemagick followed your suggestion,it indeed generated a grayscale pic of pos.jpg with a bbox,but the bbox is not located very precisely and there was no facial points aligned. The detect time is still 80s.
Do you encountered the same problem?
Thank you.

[denghuiru]

@guo253
Hi,
of course the bbox is not so perfect, because the trained model provided by the author is not very good, the trained model is only trained with 337 faces, so if you want to get a good result, you have to train a new model with larger datasets such as(AFLW). About the facial point aligned, i am also not very clear, this may wait samuel. About the detect time, I really wonder why it takes so long , the paper said it only use 30ms for a VGA image.
If you have any ideas about these problems, welcome to contact me!
Thanks!
Best regards!

[denghuiru]

@guo253 I have sent qq to you and glad to meet you!

[samuel1208]

@guo253 @denghuiru
The reason of the slow detection speed

1. The detection part is written in python, Also the trees traversing is not optimized.
2. I didn't optimize the global regression part into the trees traversing.

The reason of not show the landmark
I just use the opencv's rectgroup to merge the face rect, Actually you should use the NMS to find the best face rect instead to merge the rects. And show the landmark of te best face rect

I'll do this optimization later. If you have any advises, please tell me

[futurely]

Non maximum suppression by @melgor based on @jrosebr1's tutorial.

[melgor]

Hi. Yes, I added NMS to the demo for detection, but it return only one best rectangle.
But then I learn the model using LFWA, HELEN and IBUG and result for detection was not good (using merging or NMS). Now I am trying to get better result. If it will works, I will then send the pull-request.

[weiyt2014]

@melgor ,you use NMS，the alignment problem still be unsloved! Do you have a good idea to realise the connetion detection and alignment? hope to get suggestions!Thanks!