Processing of current captured frame and a part of previous frame simultaneously in VSDK.

Hi Mrunali,

I have pinged Vision SDK team to comment here.

Thanks,
Alex

Hi Mrunali,

 Is the intention to run face detection and eye detection algorithms in parallel threads?

Thanks and Rgds

Suriya Narayanan L

Hi Mrunali,

This is possible, i presume that you have two algorithm plugin's (one for face detection & other for the eye detection) and the output of these algorithms do not modify input frames (i.e. it operates in ALGORITHM_LINK_QUEUEMODE_NOTINPLACE mode) 

You could use a dupLink to replicate the incoming frames and pass frame simultaneously to your alog's. The face detection works on all the frames but eye detection delays it by one frame.

i.e. eye detect, on first call to process API, will save (remember the pointer to frame) frame (A) and output no eye-detected. On subsequent process API call, eye-detect algo would work on the saved frame (A) and now store the second frame (B), this is repeated for all the frames.

Please note that you will have to increase the number of buffers allocated by Link/Plugin preceeding your plugins.

Regards,

Sujith

Hi Sujith,

In Mrunali's usecase the output of face detection algorithm is fed into eye detection algorithm.

In that case, I was thinking of maintaining a shared Queue that the face detection algorithm will populate.

The eye detection algorithm can then pop out the top most element from the Queue and process it.

Thanks and Regards
Suriya Narayanan L

Hi Suriya,

In case these algo's are not required to run simultaneously. We would not require a shared queue.
We could update eye-detect algo to deffer processing of frame by one "process frame" call to achieve the required delay.

Lets wait on inputs from Mrunali.

Regards,
Sujith

Hi Supriya,

Yes my intension is to run face detection and eye detection in parallel.

Thanks
Mrunali

Hi Sujith,

Thank you for your input.
In my case, I want to delay eye detection by one frame because my eye detection takes output from face detection algorithm as its input.
The scenario is that i want to run both the algorithms in parallel. As my 2nd algorithm is dependent on the 1st, I thought of processing previous frame for eye detection.

My current framework is:

Capture -> Face_Detection_Algplugin -> Eye_Detection_Algplugin -> Display

In this case as both the algorithm works serially my FPS is very less. So as an optimization technique i thought of processing both algorithms in parallel.

Thanks,
Mrunali