Choosing specific point of detected object from box received from ssd_mobilenet_v2 retraining on custom dataset

I’m running custom object detector using retraining process of ssd_mobilenet_v2(object is a pen), retraining process have been done fine, objects are detected and bounding boxes generating. Now i need to get a specific point on the pen - middle point on the top of the pen to simulate robotic arm pick off process, currently edge detection and some image processing/filtering techniques looks encouraging but i want to know if its possible to use/retrain/inference neural network, especially on NVIDIA jetson nano/tensorrt to chose a particular point/pixel on the detected object.


In case you don’t know, we do have a library for robotic.
Please check if there is any suitable algorithm for your use case.


Thank you for response, i will try to use DOPE mentioned in Isaac SDK on jetson nano, and let know how that goes.