Good day
First, I want to thank you for this great library
Second, I want to ask about the "correlation tracker"
You mentioned that the correlation tracker is the implementation to this paper "Danelljan, Martin, et al. "Accurate scale estimation for robust visual tracking." Proceedings of the British Machine Vision Conference BMVC. 2014."
However, The paper you mentioned is a scale estimation technique only and not a tracker. It adapts MOSSE adaptive correlation tracker but, its not necessary that your implementation is MOSSE. Since MOSSE tracker works at a speed about 250 FPS, but the tracker in DLIB library able to work about 50 FPS.
Please if you tell me what kind of correlation tracker was implemented in the library because the speed difference made me wonder. If any paper about the implemented tracker is available I will be thanful to you.
If the tracker is MOSSE, why the speed about 50 FPS and not 250 FPS
I hope my quesion is clear, Thank you.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Thank you for the fast reply Mr. Davis.
Then why am I getting 50 FPS using I5 intel processor?
The video in youtube you mentioned works at 150 FPS.
Is it possible to boost the processing speed to 150. By the way I know about the AVX.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Some CPUs are slower or faster than others. Maybe this is the speed yours
runs at. Or maybe there are other things happening in your program or
computer that are taking CPU resources. Also, http://dlib.net/faq.html#Whyisdlibslow
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
One last quesion, and am sorry for burden you.
Is the source code is the same as the source code the BMVC implemented or did you wrote it yourself. I am asking that because when people ask me, I want to know reliable information
Thank you.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Good day ,
I were able to excute the code at 83 FPS using ubuntu.
On the other hand , I could not reach a 30 FPS using windows.
However, Is 83 FPS the maximum I can reach ,or can I boost it more? My processor i7-4790k 4 Ghz,
It is supporting AVX.
Last edit: yahya 2016-05-05
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Earlier on 9th of April threr wa PR #98 merged. I added possibility to provide options for corellation tracker via constructor params
If you change them, you can archive any FPS you need in the cost of tracking quality.
In my system I am not using scale tracking at all and this can gives me 2.5x performance improvement and also I am reduced filter_size param to 5
If you need to exclude scale tracking - call update_noscale() instead of update(). Its also possible to call update every 20'th frame and update_noscale - for the rest of frames
And about Windows - significantly depends on the compiler used. MinGW with gcc 5.3 onboard works well for 64 bit mode. For x86 don't forget to use correct compiler flags for better performance (-march=sandybridge -Ofast in my case)
Dlib with MinGW on Windows works faster than with MSVC
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Thank you Sir for you contribution.
I did access the code and made some modifications and the effects are good.
But it would be very helpful if you provided a link to your modified correlation tracker file and the constructor you built.
Thanks
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello! The official dlib source on github is this one https://github.com/davisking/dlib, right? I wanted to use the update_noscale() method of the correlation tracker but it is not on the correlation tracker class.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Good day
First, I want to thank you for this great library
Second, I want to ask about the "correlation tracker"
You mentioned that the correlation tracker is the implementation to this paper "Danelljan, Martin, et al. "Accurate scale estimation for robust visual tracking." Proceedings of the British Machine Vision Conference BMVC. 2014."
However, The paper you mentioned is a scale estimation technique only and not a tracker. It adapts MOSSE adaptive correlation tracker but, its not necessary that your implementation is MOSSE. Since MOSSE tracker works at a speed about 250 FPS, but the tracker in DLIB library able to work about 50 FPS.
Please if you tell me what kind of correlation tracker was implemented in the library because the speed difference made me wonder. If any paper about the implemented tracker is available I will be thanful to you.
If the tracker is MOSSE, why the speed about 50 FPS and not 250 FPS
I hope my quesion is clear, Thank you.
The BMVC paper really does describe the method, tracker included.
Thank you for the fast reply Mr. Davis.
Then why am I getting 50 FPS using I5 intel processor?
The video in youtube you mentioned works at 150 FPS.
Is it possible to boost the processing speed to 150. By the way I know about the AVX.
Some CPUs are slower or faster than others. Maybe this is the speed yours
runs at. Or maybe there are other things happening in your program or
computer that are taking CPU resources. Also,
http://dlib.net/faq.html#Whyisdlibslow
One last quesion, and am sorry for burden you.
Is the source code is the same as the source code the BMVC implemented or did you wrote it yourself. I am asking that because when people ask me, I want to know reliable information
Thank you.
I wrote the dlib code before the paper authors released their code. So the
dlib code is not based on their software.
Good day ,
I were able to excute the code at 83 FPS using ubuntu.
On the other hand , I could not reach a 30 FPS using windows.
However, Is 83 FPS the maximum I can reach ,or can I boost it more? My processor i7-4790k 4 Ghz,
It is supporting AVX.
Last edit: yahya 2016-05-05
Hello, just found your question
Earlier on 9th of April threr wa PR #98 merged. I added possibility to provide options for corellation tracker via constructor params
If you change them, you can archive any FPS you need in the cost of tracking quality.
In my system I am not using scale tracking at all and this can gives me 2.5x performance improvement and also I am reduced filter_size param to 5
If you need to exclude scale tracking - call update_noscale() instead of update(). Its also possible to call update every 20'th frame and update_noscale - for the rest of frames
And about Windows - significantly depends on the compiler used. MinGW with gcc 5.3 onboard works well for 64 bit mode. For x86 don't forget to use correct compiler flags for better performance (-march=sandybridge -Ofast in my case)
Dlib with MinGW on Windows works faster than with MSVC
Thank you Sir for you contribution.
I did access the code and made some modifications and the effects are good.
But it would be very helpful if you provided a link to your modified correlation tracker file and the constructor you built.
Thanks
He updated the official dlib source on github. So if you have the latest
dlib source code you have his updates.
Hello! The official dlib source on github is this one https://github.com/davisking/dlib, right? I wanted to use the update_noscale() method of the correlation tracker but it is not on the correlation tracker class.
Yes, that's the official source and it has the update_noscale() method on
the correlation tracker class.
Yeah, found it. Sorry for my mistake!