TDA2SX: TIDL running multiple frames

Sasa Mihajlica

Part Number: TDA2SX

Hi everyone,

I am running TIDL usecase on DSP and want to use SSD model for object detection to make inference on multiple images at sequence. NullSrc is taking images one by one and sending them for preprocessing and inference. Inference on first image is correct, but for all images after the first one inference output have 26 bboxes detected and every bbox have confidence value 0.5 and values for coordinates are not as expected for any bbox. Some of them are invalid (below 0.0 and above 1.0).
However, if TIDL link is restarted (System_linkStop, System_linkDelete, System_linkCreate, System_linkStart sequence of functions is called) after inference for one image and before another, results for next frame are correct.

I tried same usecase with Image segmentation model and using multiple input images and there wasn't this problem.

What could be the problem, why TIDL is not working correctly with my model?

Regards,
Sasa

over 6 years ago

0 Praveen Eppa1 over 6 years ago

TI__Genius 17580 points

Hi Sasa,

Did you get import tool output correct for your model ?

Thanks,

Praveen

0 Sasa Mihajlica over 6 years ago in reply to Praveen Eppa1

Intellectual 860 points

Hi Praveen,

Yes, import tool output is correct.

Regards,
Sasa

0 Sasa Mihajlica over 6 years ago in reply to Sasa Mihajlica

Intellectual 860 points

Do you have any more suggestions?
Can I at least try to have some "soft" restart for beginning? By "soft" restart I mean can I call some function at the end of AlgorithmLink_tidlProcess function to return link to initial state without taking too much time as now is the case with link restart.

Thanks,
Sasa

0 Praveen Eppa1 over 6 years ago in reply to Sasa Mihajlica

TI__Genius 17580 points

Hi Sasa,

You can try that but to root cause the issue, is it possible to run TIDL inference in standalone setup using multiple images.

Thanks,

Praveen

0 Sasa Mihajlica over 6 years ago in reply to Praveen Eppa1

Intellectual 860 points

Hi Praveen,

you mean running it using debugger to download code to DSP? If that is the case, I can't because I don't have any debugger now.
But I tried to have more frames as an input and after several frames the inference of multiple bboxes detected per frame, algorithm stabilizes itself and works OK for all of the frames.

I wonder now why it has those few frames with bad inference?.

Regards,
Sasa

0 Praveen Eppa1 over 6 years ago in reply to Sasa Mihajlica

TI__Genius 17580 points

Hi Sasa,

Thanks for the update.

You can try changing the below quantization parameters as suggested in the section 3.8 (Matching TIDL inference result) in the TIDL user guide, and check if you can see any improvement in the output.

>>Set with below parameters for running on video sequence.
createParams.quantHistoryParam1 = 20;
createParams.quantHistoryParam2 = 10;
createParams.quantMargin = 20;

Thanks,

Praveen

Processors

Processors forum

TDA2SX: TIDL running multiple frames