How can I train a keypoint detector to use with detectNet?

HI All,

I would like to add face landmarks to facenet-120 in the jetson-inference examples script for detectNet.

My issue is that the custom dataset training examples are great for training object detectors, but what I’m looking for is a custom keypoint detector. I am essentially trying to recreate RetinaFace but with my custom dataset and for use with detectNet.

I don’t want to use torch2trt just yet, because as soon as I import torch in a python script it consumes a large amount of RAM, I would love to have the model usable directly by jetson-inference’s detectNet.

Can anyone point me in the right direction to training a face/keypoint detector for use directly with detectnet?


We have a sample for landmark detector with TensorRT as well:

Since the backend inference engine are all TensorRT, you can run it with jetson_inference directly.

But please noted that the output format between detector and pose estimator is quite different. (bbox vs. keypoints)
You will need to modify the sample to handle the keypoint output.