All 16 comments
Is this a question?
davisking
on 8 Aug 2016
@davisking, no. It is a suggestion to implement YOLO in DLIB to provide a CNN based object detector. It is the natural next step over HOG+SVM.
A state-of-art combination is YOLO + SqueezeNet.
Grabber
on 9 Aug 2016
Ah, yes. There are many possible extensions. I'm adding one to dlib and it should appear in the next release.
davisking
on 9 Aug 2016
Could you provide more details? I've been working on this subject and want
to cooperate with DLIB project.
On Tuesday, 9 August 2016, Davis E. King [email protected] wrote:
Ah, yes. There are many possible extensions. I'm adding one to dlib and it
should appear in the next release.—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
https://github.com/davisking/dlib/issues/180#issuecomment-238605580, or mute
the thread
https://github.com/notifications/unsubscribe-auth/AAA9c19op8mB11HX8KHIGuPwRA_JEpsBks5qeKfcgaJpZM4Jeueg
.
Regards,
_Luiz Vitor Martinez Cardoso_
CEO & Founder at geeksys.com.br http://www.geeksys.com.br
Phone: (11) 97351-7097 | Skype: grabberbr
"If you wanna be successful, you need total dedication, go for your last
limit, give your best and love your love infinitely!"
"The only limits are the ones you place upon yourself"
Grabber
on 9 Aug 2016
The current object_detection tool in dlib is trained using this method: https://arxiv.org/abs/1502.00046. There is an obvious extension of that to the CNN framework in dlib and I'm setting that up now. What is nice about it is that it avoids any kind of hard negative mining, uses all the training windows in each image, and optimizes the actual accuracy of the detector rather than a binary classification proxy. I've already got the code running and it works well and is convenient to use in very much the same ways as the existing object_detection tooling. I'm just cleaning up the API and working on software engineering flavored usability issues at the moment. When I'm finished I'll post it to github :)
davisking
on 9 Aug 2016
@davisking how can I help your or review your code?
Grabber
on 9 Aug 2016
Not much to do right now. You can test it out when it's posted though.
That is always helpful :)
davisking
on 9 Aug 2016
When are you planning to?
On Tuesday, 9 August 2016, Davis E. King [email protected] wrote:
Not much to do right now. You can test it out when it's posted though.
That is always helpful :)—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
https://github.com/davisking/dlib/issues/180#issuecomment-238688819, or mute
the thread
https://github.com/notifications/unsubscribe-auth/AAA9c_PnE1p6NEVO0Cbuvj9QThsvZZ7Uks5qeOm8gaJpZM4Jeueg
Grabber
on 9 Aug 2016
I'm not sure. Hopefully in two weeks.
davisking
on 10 Aug 2016
Have you tried it with SqueezeNet?
Grabber
on 10 Aug 2016
no
davisking
on 10 Aug 2016
@davisking have you seen https://github.com/antikantian/facedet-squeezenet ?
Grabber
on 11 Aug 2016
No. Nice to see more people using dlib's new API though :)
davisking
on 11 Aug 2016
@davisking any update on this feature?
Grabber
on 3 Sep 2016
It's all working. I'll probably post the code and an example this weekend.
davisking
on 3 Sep 2016
Check this out :) https://github.com/davisking/dlib/blob/master/examples/dnn_mmod_ex.cpp
davisking
on 5 Sep 2016
Related issues
yourmailhacked
·
3Comments
abhisheksoni27
·
4Comments
reunanen
·
3Comments
pliablepixels
·
4Comments
alison-carrera-pegasus
·
3Comments
Most helpful comment
The current object_detection tool in dlib is trained using this method: https://arxiv.org/abs/1502.00046. There is an obvious extension of that to the CNN framework in dlib and I'm setting that up now. What is nice about it is that it avoids any kind of hard negative mining, uses all the training windows in each image, and optimizes the actual accuracy of the detector rather than a binary classification proxy. I've already got the code running and it works well and is convenient to use in very much the same ways as the existing object_detection tooling. I'm just cleaning up the API and working on software engineering flavored usability issues at the moment. When I'm finished I'll post it to github :)