ocl_samples_facedetect fails on Intel platform with FEATURE_CL_INTEL_DEVICE (Bug #3566)


Added by xiuli pan about 11 years ago. Updated about 11 years ago.


Status:Done Start date:2014-02-24
Priority:Normal Due date:
Assignee:Alexander Karsakov % Done:

0%

Category:ocl
Target version:2.4.9
Affected version:2.4.8 (latest release) Operating System:Linux
Difficulty: HW Platform:x64
Pull request:https://github.com/Itseez/opencv/pull/2421

Description

The ocl example facedetect based on haar features returns no correct results when using the Intel platform with both Windows and Linux OpenCL, but can work fine on Nvidia or AMD OpenCL. Then if I comment all the codes related with FEATURE_CL_INTEL_DEVICE in /modules/ocl/src/haar.cpp the sample can still work fine on Intel platform.
(See related pull request 1675)


compare_HD4600_vs_CPU.jpg (200.9 kB) Alexander Karsakov, 2014-02-26 09:44 am


Associated revisions

Revision 2ecca8f5
Added by Vadim Pisarevsky about 10 years ago

Merge pull request #3566 from jet47:cuda-imgproc-refactoring

History

Updated by Alexander Karsakov about 11 years ago

  • Status changed from New to Open
  • Target version changed from Next Hackathon to 2.4.9
  • Assignee set to Andrey Pavlenko

Updated by Alexander Karsakov about 11 years ago

  • Category set to ocl

Updated by Andrey Pavlenko about 11 years ago

Alexander, this is Intel-specific issue.

  • Assignee changed from Andrey Pavlenko to Alexander Karsakov

Updated by Andrey Pavlenko about 11 years ago

  • Description changed from The ocl example facedetect based on haar features returns no correct results ... to The ocl example facedetect based on haar features returns no correct results ... More

Updated by Andrey Pavlenko about 11 years ago

  • Description changed from The ocl example facedetect based on haar features returns no correct results ... to The ocl example facedetect based on haar features returns no correct results ... More

Updated by Alexander Karsakov about 11 years ago

Hi xiuli pan,
Could you please provide images which classifier failed to face detection? It will be very useful for solving this problem.

Thanks,
Alexander.

Updated by xiuli pan about 11 years ago

Here is the pictures. Also, the candidate is about 220,000 for the lena.jpg when I uncomment the code //printf("candidate[0]=%d\n",candidate["\0]);, that also makes the detect really slow.


(The picture sever I use may be broken. If you can not see the picture please email me )

Updated by Alexander Karsakov about 11 years ago

I can't exactly reproduce problem with images you sent. What settings do you use to obtain this result?

Updated by xiuli pan about 11 years ago

I am using the 3rd Intel Graphic HD4000, maybe that's the cause? I see the update is after the Intel release the OpenCL sdk for 4rd Intel Graphics.

Updated by Alexander Karsakov about 11 years ago

Could you please try to use patch from this pull request - https://github.com/Itseez/opencv/pull/2421. This patch works for me.

I am using the 3rd Intel Graphic HD4000, maybe that's the cause? I see the update is after the Intel release the OpenCL sdk for 4rd Intel Graphics.

Yes, this may be the cause.
I tested with latest Intel OpenCL SDK on Windows. Also I suggest you to update driver for graphics adapter.

  • Pull request set to https://github.com/Itseez/opencv/pull/2421

Updated by xiuli pan about 11 years ago

I have tried the patch you provide and it seems can not find the right face still. I will test on more platform tommorrow.

Updated by xiuli pan about 11 years ago

It seems the patch can work fine on the windows of HD4000, but still fail on Linux.

Updated by Alexander Karsakov about 11 years ago

OK. I will try to check it on Linux.

Updated by xiuli pan about 11 years ago

I have compared the output between the Windows and Linux, and find a bug of the Linux sdk. Thank you for your help. And the patch for the OCL you provide is very useful, it speed up the detection. Should this report be cancelled or complete?

Updated by Alexander Karsakov about 11 years ago

I think this report should be completed.
What version of Linux SDK do you use?

  • Status changed from Open to Done

Also available in: Atom PDF