Person arms in the Jetson Inference object detector

[usuario1918]

I remember that I did it in the past but I don’t remember how. I want the people detector to only detect the body of people without arms, that is, the rectangle does not get bigger when I extend my arms.

How can I do it?

from

jetson_inference import detectNet
net = detectNet("ssd-mobilenet-v2", threshold=0.6)

while True:
    img = input.Capture()
    detections = net.Detect(img, overlay='box') # box,labels,conf or none

[AastaLLL]

Hi,

Do you do this with a detector or a pose estimator?
For the detector, maybe you can consider the aspect ratio of the bounding box.

Thanks.

[usuario1918]

@usuario1918

net = detectNet("ssd-mobilenet-v2", threshold=0.6)

I do it with a detector, I placed the model. The pose estimator goes to 14 FPS and the detector to 17 FPS, that’s why the object detector suits me.

@AastaLLL

For the detector, maybe you can consider the aspect ratio of the bounding box.

I don’t understand this, clarify it for me, please. Like I said, I remember getting it done before but I forgot what that line of code looks like. That is, when I extended my arms, the box did not increase.

[dusty_nv]

You could attempt to edit the bounding boxes by running detections = net.Detect(img, overlay='none'), then modifying the detections (i.e. by aspect ratio or some other metric you define), and then run net.Overlay(img, detections) to generate the bounding-box overlay with your updated detections. However, this may not be robust in practice and you could either use the pose model as @AastaLLL suggested, or train your own detection model on a dataset which did not include the arms. Although you also seem to have a solution from before, I’m not sure what that was however as “no arms” is not a feature supported in jetson_inference.detectNet