RTOS/TDA3XEVM: How use the ext PD_Weight.bin for ObjectDetection model

user5110132

Expert 1950 points

Part Number: TDA3XEVM

Tool/software: TI-RTOS

I follow the ObjectDetection_DSP_UserGuide.pdf to use my own train model for PD detection?

1. I set the inArgs.pdConfig.useExtWeightsPD = TRUE

2. I put the date in the buffer of pInBufs->bufDesc[TI_OD_IN_BUFDESC_PD_ADABOOST_WEIGHTS]->bufPlanes[0].buf = pObj->pWeights;

3. And I run the demo, the result is weird as follow, some date are not right and the output of vedio is bule, maybe some mem error

[DSP1 ] 20.884615 s: ALG_OBJECT_DETECT: Objects = 1 !!!
[DSP1 ] 20.884676 s: ALG_OBJECT_DETECT: #0: Tag=0, Type=0, SubType=0, Score=0.000000, Scale=0.000000, (0, 0 : 0 x 0)
[DSP1 ] 21.811718 s: ALG_OBJECT_DETECT: Objects = 3 !!!
[DSP1 ] 21.811779 s: ALG_OBJECT_DETECT: #0: Tag=72, Type=0, SubType=0, Score=82.137421, Scale=1.000000, (38, 12 : 24 x 56)
[DSP1 ] 21.811840 s: ALG_OBJECT_DETECT: #1: Tag=0, Type=0, SubType=0, Score=0.000000, Scale=0.000000, (0, 0 : 0 x 0)
[DSP1 ] 21.811901 s: ALG_OBJECT_DETECT: #2: Tag=0, Type=0, SubType=0, Score=0.000000, Scale=0.000000, (0, 0 : 0 x 0)
[DSP1 ] 21.947325 s: ALG_OBJECT_DETECT: Objects = 2 !!!
[DSP1 ] 21.947386 s: ALG_OBJECT_DETECT: #0: Tag=1769495, Type=786450, SubType=1245197, Score=0.000000, Scale=0.000000, (393234, 22282482 : 2490533 x 6488107)
[DSP1 ] 21.947447 s: ALG_OBJECT_DETECT: #1: Tag=26738927, Type=7995668, SubType=3997756, Score=0.000000, Scale=0.000000, (3997775, 1769515 : 3080233 x 3407933)
[DSP1 ] 30.660983 s: ALG_OBJECT_DETECT: Objects = 1 !!!
[DSP1 ] 30.661044 s: ALG_OBJECT_DETECT: #0: Tag=0, Type=0, SubType=0, Score=0.000000, Scale=0.000000, (0, 0 : 0 x 0)
[DSP1 ] 31.460684 s: ALG_OBJECT_DETECT: Objects = 1 !!!
[DSP1 ] 31.460714 s: ALG_OBJECT_DETECT: #0: Tag=0, Type=0, SubType=0, Score=0.000000, Scale=0.000000, (0, 0 : 0 x 0)
[DSP1 ] 32.269627 s: ALG_OBJECT_DETECT: Objects = 1 !!!
[DSP1 ] 32.269688 s: ALG_OBJECT_DETECT: #0: Tag=0, Type=0, SubType=0, Score=0.000000, Scale=0.000000, (0, 0 : 0 x 0)
[DSP1 ] 33.073415 s: ALG_OBJECT_DETECT: Objects = 1 !!!
[DSP1 ] 33.073445 s: ALG_OBJECT_DETECT: #0: Tag=0, Type=0, SubType=0, Score=0.000000, Scale=0.000000, (0, 0 : 0 x 0)
[DSP1 ] 33.209418 s: ALG_OBJECT_DETECT: Objects = 3 !!!
[DSP1 ] 33.209479 s: ALG_OBJECT_DETECT: #0: Tag=72, Type=0, SubType=0, Score=82.137421, Scale=1.000000, (38, 12 : 24 x 56)
[DSP1 ] 33.209540 s: ALG_OBJECT_DETECT: #1: Tag=0, Type=0, SubType=0, Score=0.000000, Scale=0.000000, (0, 0 : 0 x 0)
[DSP1 ] 33.209601 s: ALG_OBJECT_DETECT: #2: Tag=0, Type=0, SubType=0, Score=0.000000, Scale=0.000000, (0, 0 : 0 x 0)
[DSP1 ] 33.343286 s: ALG_OBJECT_DETECT: Objects = 1 !!!
[DSP1 ] 33.343347 s: ALG_OBJECT_DETECT: #0: Tag=1769495, Type=786450, SubType=1245197, Score=0.000000, Scale=0.000000, (393234, 22282482 : 2490533 x 6488107)
[DSP1 ] 34.690538 s: ALG_OBJECT_DETECT: Objects = 1 !!!
[DSP1 ] 34.690569 s: ALG_OBJECT_DETECT: #0: Tag=0, Type=0, SubType=0, Score=0.000000, Scale=0.000000, (0, 0 : 0 x 0)
[DSP1 ] 35.767217 s: ALG_OBJECT_DETECT: Objects = 1 !!!
[DSP1 ] 35.767248 s: ALG_OBJECT_DETECT: #0: Tag=1769495, Type=786450, SubType=1245197, Score=0.000000, Scale=0.000000, (393234, 22282482 : 2490533 x 6488107)
[DSP1 ] 36.033489 s: ALG_OBJECT_DETECT: Objects = 1 !!!
[DSP1 ] 36.033520 s: ALG_OBJECT_DETECT: #0: Tag=72, Type=0, SubType=0, Score=82.137421, Scale=1.000000, (38, 12 : 24 x 56)
[DSP1 ] 36.305008 s: ALG_OBJECT_DETECT: Objects = 1 !!!
[DSP1 ] 36.305038 s: ALG_OBJECT_DETECT: #0: Tag=0, Type=0, SubType=0, Score=0.000000, Scale=0.000000, (0, 0 : 0 x 0)

AND the UserGuide of Object detection is so simple than I don't know what's mean of Tag=72?

Waiting for your help

over 7 years ago

0 Pramod Kumar Swami over 7 years ago

TI__Genius 14100 points

Are you using trained model with same tree structure as TI's model? The OD can accept esternal trained model but not of any arbitrary structure.
Also did you trained using ACF-jacinto training framework?
github.com/.../acf-jacinto

Thanks,
With Regards,
Pramod

0 Shyam Jagannathan over 7 years ago

TI__Genius 10365 points

Tag is used for uniquely identifying objects across frames. This comes in handly during object tracking. Dont worry about this field in general. Can you please share some details on your AdaBoost classifier structure? how many trees? tree structure? model size etc..

0 user5110132 over 7 years ago in reply to Pramod Kumar Swami

Expert 1950 points

Yes, I using the ACF-jacinto training framework; But I just using the 1024 trees, 3 level, the model size is [64, 32];
It that means I could not change any train parameter??

0 user5110132 over 7 years ago in reply to Shyam Jagannathan

Expert 1950 points

OK I will show the detector ops for you using picture

0 user5110132 over 7 years ago in reply to user5110132

Expert 1950 points

0 user5110132 over 7 years ago in reply to Pramod Kumar Swami

Expert 1950 points

What is the eternal train model support? Is that the acf-jacinto default parameter?
But the default parameter of cellsize is 8 in train model. But in the demo of TI uscecase, the cellsize is 4, It so weire!

0 Shyam Jagannathan over 7 years ago in reply to user5110132

TI__Genius 10365 points

For TI uscase the cell size is also 8x8. EVE is used to create 4x4 cells and then DSP does a 2x2 of 4x4 cells on-the-fly to compute 8x8 cells.
Can you please provide, conifg parameters to EVE and C66x DSP as well?

You can have model size of 64x32 and any number of trees but adaboost tree structure should not be more than 3 nodes and 4 leaves

0 user5110132 over 7 years ago in reply to Shyam Jagannathan

Expert 1950 points

This is my config parameters to EVE, I don't Change the parameter of computing Hog feature in EVE

My config parameters to execute the OD module is like this

Did the ROI of this bufPlanes need to set???

Thank you!

Waiting for your help!!

0 Shyam Jagannathan over 7 years ago in reply to user5110132

TI__Genius 10365 points

I see that you have commented out TI_OD_IN_BUFDESC_FEATURE_PLANES. This is the output of EVE (feature pyramid) and input to OD module. This has to be provided "as-is". The buffer passed from EVE to DSP not only contains the feature pyramid data but also some headers which DSP requires to understand the layout of feature pyramid. So if you are not changing EVE code, then please pass this information as-is.

Regards,
Shyam

0 user5110132 over 7 years ago in reply to user5110132

Expert 1950 points

But it still not work!

0 Shyam Jagannathan over 7 years ago in reply to user5110132

TI__Genius 10365 points

1. Can you please confirm that default PD,VD,TSR demo is working fine?
2 . Can you please confirm if the descriptor file generated by ACF Jacinto is providing weights in fixed-point format and not floating-point?
3. Can you please confirm that you are using conversion utils "ti_object_detection\utils\AdaboostTableGen.exe" to convert from ACF Jacinto descriptor to a binary file something like "vd_adaboost_weights.bin"?

If you have done all this correctly, then you should see the algorithm pick up new weights and classify correctly.

0 user5110132 over 7 years ago in reply to Shyam Jagannathan

Expert 1950 points

Would the feature pyramid scale and step influence the result？

0 user5110132 over 7 years ago in reply to Shyam Jagannathan

Expert 1950 points

1. Yes ,The PD demo is working fine, But when I used the external pd_adaboost_weights.bin which I train it to detect the person in video. It is not ok?

2. Yes It is fixed-point , I can give you the files;

3. Yes

0 user5110132 over 7 years ago in reply to Shyam Jagannathan

Expert 1950 points

But How can I give you files

0 user5110132 over 7 years ago in reply to Shyam Jagannathan

Expert 1950 points

How can Set the Hog_Feature Input Buf

This is the C:\PROCESSOR_SDK_VISION_03_02_00_00\ti_components\algorithms\REL.200.V.OD.C66X.00.06.02.00\200.V.OD.C66X.00.06\modules\ti_object_detection\test\src\object_detection_tb.c

The height is 1, and the width is the 10*1024*1024

and The follow is the Alg_ObjectDetect the width is the image width; and the height is 10*height.

Make me confuse, I will follow which to coding

0 user5110132 over 7 years ago in reply to Shyam Jagannathan

Expert 1950 points

I think the problem is that how can I put the pd_adaboost_weight put the right memory! Is this some userguider for programmer. The OD module UserGuider is to simple that can't give us a guider

0 Shyam Jagannathan over 7 years ago in reply to user5110132

TI__Genius 10365 points

Shiju,

Can you please help the user create a buffer to provide external weights to OD module?
This information will not be part of the OD algorithm as this is a PSDK-Vision concept.

Regards,
Shyam

0 Shyam Jagannathan over 7 years ago in reply to Shyam Jagannathan

TI__Genius 10365 points

PSDK-Vision uses default weights which is built-in part of the OD libraray. If you have to supply weights externally please take a look at ti_components\algorithms\REL.200.V.OD.C66X.00.06.02.00\200.V.OD.C66X.00.06\modules\ti_object_detection\test\src\object_detection_tb.c

You will basically have to allocate additional buffers in PSDK-Vision to enable this feature,
File to modify (just like how its done in test bench) is vision_sdk\apps\src\rtos\alg_plugins\objectdetection\objectDetectionLink_algPlugin.c

0 user5110132 over 7 years ago in reply to Shyam Jagannathan

Expert 1950 points

I do all the step as your said; I use the pd_adaboost_weights.bin provided in ti_components\algorithms\REL.200.V.OD.C66X.00.06.02.00\200.V.OD.C66X.00.06\modules\ti_object_detection\test\ as external model input . The demo of Object Detection CAN run well.

BUT When replace the pd_adaboost_weights.bin which was generated by the tool AdaboostTableGen.exe. The input file is AcfJacintoInriaDetector.descriptor from the TI ACF jacinto source code. I don't change any information.
It is not work..

0 Luke Lin over 7 years ago in reply to user5110132

TI__Expert 3250 points

Hi,

If you check the sample "vehicle.descriptor" included in the OD, you will find below two comment lines in the beginning.

#FeatureScalingLambda: -0.000000 -0.000000 -0.000000 -0.000000 -0.000000 -0.000000 -0.000000 -0.000000 -0.000000 -0.000000
#AdaboostCascadeDescriptor: 1280 2 13 7

Currently the descriptor generated by ACF Jacinto has only one comment line in the beginning.

#AdaboostCascadeDescriptor: 1280 2 13 7

AdaboostTableGen.exe skips two comment lines then starts conversion. This will cause error if the descriptor has only one comment line.

Please try adding an arbitrarily comment line in the beginning of your descriptor, convert it to .bin, see if it works.

0 Shyam Jagannathan over 7 years ago in reply to Luke Lin

TI__Genius 10365 points

No this is not the right way. Those are not the only two lines that are extra. Please compare vehicle.descriptor file and AcfJacintoInriaDetector.descriptor file. The vehicle.descriptor file is written in "Cascade" format whereas the AcfJacintoInriaDetector.descriptor file is written in "Compact" format.

The right way is to provide a descriptor file written in "Cascade" format to AdaboostTableGen.exe.

Regards,
Shyam

0 user5110132 over 7 years ago in reply to Luke Lin

Expert 1950 points

OK I can try it! I hope It can work!!

0 user5110132 over 7 years ago in reply to Luke Lin

Expert 1950 points

Thank you! Luke Lin!

As your suggesting ! This problem was solved! But This tips should be wrote in user guider. This bug is so difficulty to fix. If you don't tell me. I can not never find it!

Processors

Processors forum

RTOS/TDA3XEVM: How use the ext PD_Weight.bin for ObjectDetection model