Single Model for Person and Face other then PeopleNet

[devesh]

• Hardware Platform (Jetson / GPU) Jetson Nano 2GB
• DeepStream Version 5.1
• JetPack Version (valid for Jetson only) 4.5.1 [L4T 32.5.1]
• TensorRT Version 7.1.3.0
• Issue Type( questions, new requirements, bugs) question

Hi,
I am trying to create a deepstream application for realtime inference over people and face.Since I have used two different models for person detection(Resnet10.caffemodel) and then face detection(fd_lpd.caffemodel taken from here), so with this approach i am getting pretty low FPS.

So, I am looking for single model which can detect both classes.

Any help would be appreciated.
Thanks.

[kayccc]

Hi devesh.tiwari444,

Sorry for the late response, have you managed to get issue resolved?

Thanks

[devesh]

Hi @kayccc ,
I am still looking for such model which can behave as primary detector and can detect person and their face as well.
Looking forward for your suggestions.
Thanks.

[Morganh]

Hi @devesh
Peoplenet can detect person and face classes. Do you have a chance to try it?
See Overview — TAO Toolkit 3.22.05 documentation
and https://ngc.nvidia.com/catalog/models/nvidia:tlt_peoplenet

[devesh]

Hi @Morganh
Thanks for your response, can you please suggest me a TRT based model.
Currently i have some limitations for using only TRT models.
Thanks.

[Morganh]

Actually deepstream can deploy the etlt model.See DetectNet_v2 — Transfer Learning Toolkit 3.0 documentation

If you need a tensorrt engine, there is a tool named tlt-converter. It can convert the etlt model to trt engine. You can copy the etlt model to the machine where you want to run inference. Refer to DetectNet_v2 — Transfer Learning Toolkit 3.0 documentation

[devesh]

Hi @Morganh,
Thanks for your help given links were helpful and they solved my issue, just one last question before closing the topic.
Accuracy for face detection in peoplenet is quit low can you please suggest me to what should I do to get Peoplenet more accurate on face.
Thanks.

[Morganh]

End users can retrain peoplenet with more data of face class. As mentioned in https://ngc.nvidia.com/catalog/models/nvidia:tlt_peoplenet,

Although bag and face class are included in the model, the accuracy of these classes will be much lower than people class. Some re-training will be required on these classes to improve accuracy.

BTW, moving this topic from DS forum into TLT forum.