May I say that the answer is within your message? When you apply YOLO or any other related algorithms on still images, then you end-up with a static box (I am stating the obvious here). However, in live streams you are dealing with successive frames and therefore, either the ROIs (i.e., objects) are in motion or the camera is in motion or both*. Hence the YOLO algorithm will be tracking the ROI which you are targeting and the box boundaries (e.g., confidence) will slightly change producing the flickering.

-----------

(*) Unless your live stream is from a fixed camera on a fixed object (still other interference -like lighting conditions- might trigger the flickering.

Dear Sahil Dalal

Basic hurdles that object detection are facing:

1. Variation in viewpoint. One of the most difficult aspects of object detection is that an object can appear completely different when viewed from different angles.

2. Deformation.

3. Occlusion.

4. Lighting conditions.

5. Background that is cluttered or textured.

6. Variety.

7. Speed.

For more refer: https://www.exposit.com/blog/computer-vision-object-detection-challenges-faced/

Kind Regards

Qamar Ul Islam

Abbas Cheddad and Qamar Ul Islam

Thanks for the response sir

Your points are right. But, situation is quite different here.

There are two situations here:

1. Running the test application of the EVK Board, with image sensor for capturing live stream. Here, flickering issue is coming.

2. Then, I tried to input the live streaming from EVK board (using rtsp) to the python code of mine in my laptop. In that case, it is working fine. No flickering of bounding box.

In both these cases, same YOLOv5s model is used.

I think, it is some issue with the code of the test application.

Please suggest anything which is useful in this situation.

Thanks in advance.

Sahil Dalal

Then the problem might be arising at the acquisition stage using the sensor and the deployed program on the EVK Board (e.g., electronics or test program issue or both). In this case, my knowledge is very limited unfortunately.

But if you can get the input of the live streaming to your laptop without flickering, which is what you will be using for further processing and analysis of the stream-data, then why don't you just proceed?

I recently supervised a Master thesis with the title "Stabilizing Non-Maximum-Suppression: More Stable Replacement for Non-Maximum-Suppression in Object and Face Detectors" where the author has pointed out that the Non-Maximum Suppression (NMS) algorithm applied in all object and face detectors is the problem that makes bounding boxes jitter substantially.

He found that one algorithm is able to improve this, which is:

Bodla, N., Singh, B., Chellappa, R., and Davis, L. S.: Improving Object Detection With One Line of Code. ICCV 2017

Another algorithm proposed by himself was to take a weighted average of the detected bounding boxes instead of suppressing overlapping boxes. Incorporating such a different version of NMS into object detectors is not that easy, though,

Unfortunately, the thesis is not yet published, but I can send you a version of that privately.

Abbas Cheddad

Yes sir, I am getting your point.

Flickering must be because of some parameter in the program.

And sir, about the code running on my laptop with live streaming from board, that I can't use because we are checking the working of the test applications as they are already given with the board. I am trying to resolve the errors in them first. If that doesn't work well, then last option will be my code.

But Thank you for your response sir.

Thanks a lot.

Manuel Günther

Thank you for the response sir

But I think, bounding boxes jitter is something different. Its like vibration or shaking in the bounding box.

In my question, I am facing flickering of the bounding box. Flickering of the bounding box while detecting objects is like blinking of eyes means for just microsecond bounding box appears and then it disappears. After 3-4 seconds, it again appears for a microsecond and then again disappears.

Still sir, may be I am wrong, the dissertation about which you are talking may be for the same issue that I am facing. I really want to look that thesis once. If you can share that please send me the thesis.

Thank you.