TDA2: SSD different result between quantStatusTool and TDA2EXVEM

Lee...

Part Number: TDA2

Hi:

I traine a SSD model changed on "github.com/tidsp/caffe-jacinto-models/tree/caffe-0.16/trained/object_detection/voc0712".

With quantStatusTool,the result is good,almost every target object is detected,But when I run the model on TDA2EXEVM,some object could't be detected in some frames.for example,as below

this result is runned with quantStatusTool;but the "1" object couldn't be detected with TIDL_OD usecase when runned on TDA2EXEVM.

And with quantStatusTool,I found the "1" object's confidence is 0.316087,so I set the "scoreThreshold" to 0.3 in TIDL_OD usecase,but the "1" object still couldn't be detected.

Is there something I missed?

My import config file is below:

     # Default - 0
randParams         = 0

# 0: Caffe, 1: TensorFlow, Default - 0
modelType          = 0

# 0: Fixed quantization By tarininng Framework, 1: Dyanamic quantization by TIDL, Default - 1
quantizationStyle = 1

# quantRoundAdd/100 will be added while rounding to integer, Default - 50
quantRoundAdd      = 25

numParamBits       = 8
# 0 : 8bit Unsigned, 1 : 8bit Signed Default - 1
inElementType      = 0

inputNetFile       = "..\..\test\testvecs\config\caffe_modesl\ssd480x192\sparse\deploy.prototxt
inputParamsFile       = "..\..\test\testvecs\config\caffe_modesl\ssd480x192\sparse\ssd480x192_sparse_iter_100000.caffemodel"
outputNetFile      = "..\..\test\testvecs\config\tidl_models\tidl_net_jdetNet_480x192_ssd.bin"
outputParamsFile   = "..\..\test\testvecs\config\tidl_models\tidl_param_jdetNet_480x192_ssd.bin"

rawSampleInData = 1
preProcType   = 4
sampleInData = "..\..\test\testvecs\input\trace_480x192.y"
tidlStatsTool = "..\quantStatsTool\eve_test_dl_algo.out.exe"
layersGroupId = 0   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   2   1   2   1   2   1   2   1   2   1   2   1   2   1   2   1   2   1   2   2   2   2   2   2   2   0
conv2dKernelType = 0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1
#conv2dKernelType = 1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1   1

Regards

over 5 years ago

0 Praveen Eppa1 over 5 years ago

TI__Genius 17580 points

Hi Lee,
Can you please share the caffe model and prototxt to replicate the issue at our end?

Thanks,
Praveen

0 Lee... over 5 years ago in reply to Praveen Eppa1

Intellectual 410 points

Hi:

there is my caffe model and prototxt.My model only detect 1 calss:car.

models&prototxt.7z

Looking forward to your reply.

Regards

0 Praveen Eppa1 over 5 years ago in reply to Lee...

TI__Genius 17580 points

Hi,

Can you please also share the input image that you used to replicate the issue quickly?

Thanks,
Praveen

0 Lee... over 5 years ago in reply to Praveen Eppa1

Intellectual 410 points

Hi:

there is some images from my test video;I test part of them:In most of images ,the object could be detected with quantStatusTool,but couldn't be on on TDA2EXEVM with TIDL_OD.for example,the image "cqh20.jpg".

test imgs.7z

0 Praveen Eppa1 over 5 years ago in reply to Lee...

TI__Genius 17580 points

Hi,

Thanks for sharing the input.

I could replicate the issue. I will debug and send you the fix.

Thanks,
Praveen

0 Praveen Eppa1 over 5 years ago in reply to Praveen Eppa1

TI__Genius 17580 points

Hi Lee,

Did you run single image (or frame) on TIDL use case and comparing the output with import tool output and detentions are not matching ?

or

If you running a video and output (some dectections) are not matching with import tool detentions, then this is expected as there could be some issue with quantization and this output will come to stable after some time.
Please refer to section 3.8 Matching TIDL inference result in the user guide on how to set the quantization parameters for a video sequence.

Thanks,
Praveen

0 Lee... over 5 years ago in reply to Praveen Eppa1

Intellectual 410 points

Hi:

I run a video;and I tryed the "quantHistoryParam1"、"quantHistoryParam2" and "quantMargin" with several different values,like

"quantHistoryParam1 = 20(or 0,10);

quantHistoryParam2 = 10(or 0,5) ;

quantMargin = 20(or 0,10)"

But result is stille same:in some frame of input video,the "1" object can't be detected.

What value shoule I set for the the "quantHistoryParam1"、"quantHistoryParam2" and "quantMargin"?

there is a clip of my test video data and header:

data&header.7z

Regards

0 Praveen Eppa1 over 5 years ago in reply to Lee...

TI__Genius 17580 points

Hi Lee,

To debug this issue, first we need to replicate it on the single image instead of video.

Could please first run the input image (used to import the models) on the EVM using standalone and check the issue is observed on not?

Please refer to section 3.3.5 in the user guide to run on the standalone.

Thanks,
Praveen

0 Lee... over 5 years ago in reply to Praveen Eppa1

Intellectual 410 points

sorry I don't have a "USB Emulator";I only can work as "Load using SD card".

0 Praveen Eppa1 over 5 years ago in reply to Lee...

TI__Genius 17580 points

Hi Lee,

Okay, I will try to replicate the issue in VSDK at our end and will let you know.

Thanks,
Praveen

0 Praveen Eppa1 over 5 years ago in reply to Praveen Eppa1

TI__Genius 17580 points

Hi Lee,

There is no problem in the code for this issue.
This issue may be because of quantization , for this kind of issues, use first frame of the video for quant stats collection in the import and then use those for inference, first frame detection's will match exactly and will take some time for the rest of the frames for quantization to stabilise.

Thanks,
Praveen

0 Lee... over 5 years ago in reply to Praveen Eppa1

Intellectual 410 points

Hi:
Thanks very much.
But then what could I do to resolve this object detection lost problem?

Regards

0 Praveen Eppa1 over 5 years ago in reply to Lee...

TI__Genius 17580 points

Hi,

Did you try matching the first frame that you imported?

I would recommend you to try our training model shared on github , the link is below..
github.com/.../JDetNet

Thanks,
Praveen

Processors

Processors forum

TDA2: SSD different result between quantStatusTool and TDA2EXVEM