Overlapping Bounding Box Coordinates

[faiz26]

Hi,

I am trying to implement NMS (Non-Maxima Suppression) to solve the issue of multiple and overlapping bounding box when detecting an object. But, the NMS function that I am using require a list of bounding boxes as an input.

This is the NMS function that I am referring to,

PyImageSearch/imutils/blob/master/imutils/object_detection.py

However, the output that that I obtain from net.Detect are as shown below, which are not in the format that I can feed into the NMS function.

So, my questions are:

• How can I get all of the overlapping bounding boxes in a “LIST”? So that I can feed into the NMS function and ensure it works as expected.
• Is there any other method to apply NMS? Maybe via Jetson Inference Package or something?

Thank you.

[AastaLLL]

Hi,

detectNet do implement a non-maximal suppression itself.

dusty-nv/jetson-inference/blob/master/c/detectNet.cpp#L813

To get the whole overlapping bbox, you can set numDetections += 1; to following line:
https://github.com/dusty-nv/jetson-inference/blob/master/c/detectNet.cpp#L856

This indicates always add the detections to the array without suppression.

Thanks.

[faiz26]

I am sorry, but this a sample from my detection output. I am using a custom model and not running the detection model via a docker container.

In the picture below are a few lines of my codes for this application. Line 44 is where I called the net.Detect function.

[dusty_nv]

@faiz26

I am sorry, but this a sample from my detection output. I am using a custom model and not running the detection model via a docker container.

You may want to decrease the clustering overlap threshold on this line of code:

https://github.com/dusty-nv/jetson-inference/blob/2fb798e3e4895b51ce7315826297cf321f4bd577/c/detectNet.h#L477

Then re-run make && sudo make install

@faiz26

How can I get all of the overlapping bounding boxes in a “LIST”? So that I can feed into the NMS function and ensure it works as expected.

You can use list comprehension to make it:

box_list = [(det.Left, det.Top, det.Right, det.Bottom) for det in detections]

(this is assuming your NMS function also expects the bounding box in (Left, Top, Right, Bottom) coordinate order)

[faiz26]

Hi dusty,

Thank you so much! both of your recommendation works!

The easiest solution is by decreasing the clustering overlap threshold. I lowered the value down to 0.5 and the results are very good for me. TRUST ME, Applying another/external NMS might work at a certain level, but it is much troublesome to reach a stage just to ensure it work as expected.