(DeepStream on NX) How to use facenet-120 with nvinfer and nvdsosd without custom bounding box parsing?

[excubiteur]

This is a follow up to:

Is there a face detection model that is known to work with Xavier NX Jetpack 4.4 GA TRT

This is related to Xavier NX Jetpack 4.4 GA TRT gives wrong results for a specific caffe model. On Xavier NX Jetpack 4.4 GA TRT, res10_300x300_ssd_iter_140000_fp16.caffemodel does not work. Is there a face detection model that is known to work with Xavier NX Jetpack 4.4 GA TRT?

Referring to

GitHub - excubiteur/deepstream.facenet

I have a deepstream pipeline using facenet with nvinfer and nvdsosd without custom bounding box parsing:

excubiteur/deepstream.facenet/blob/master/launch.sh

nvinfer uses this config:

excubiteur/deepstream.facenet/blob/master/config_facenet_primary.txt

When I run the pipeline, the video output is not showing any bounding boxes over faces in the video.

[AastaLLL]

Hi,

Facenet is based on the DetectNet architecture which is not supported in the Deepstream bbox parser.
Please update a custom parser based on this document:
https://docs.nvidia.com/metropolis/deepstream/dev-guide/index.html#page/DeepStream_Development_Guide/deepstream_custom_model.html#wwpID0ELHA

Ex.

parse-bbox-func-name=[function/name]
custom-lib-path=[library/path/xxx.so]

The implementation to convert output tensor into bounding box of detectnet can be found here:

dusty-nv/jetson-inference/blob/master/c/detectNet.cpp#L834

Thanks.