Training and Optimizing a 2D Pose Estimation Model with the NVIDIA Transfer Learning Toolkit, Part 1

[jwitsoe]

Originally published at: https://developer.nvidia.com/blog/training-optimizing-2d-pose-estimation-model-with-tlt-part-1/

Human pose estimation is a popular computer vision task of estimating key points on a person’s body such as eyes, arms, and legs. This can help classify a person’s actions, such as standing, sitting, walking, lying down, jumping, and so on. Understanding the context of what a person might be doing in a scene has…

[MingGatsby]

The model of the training , it can be used at the following article?

Creating a Human Pose Estimation Application with NVIDIA DeepStream

Some additional info. I noticed when starting the app with different models, the number of layers are different and it seems that the order of the dimension info is different between the pose model and the resnet/densenet models I converted myself. Since the expected order is hard coded in find_peaks() and refine_peaks() of post_process.cpp, the outcome is different between different models. I tried correcting this, but it still doesn’t give me good results. So the question is if I somehow m…

I know I can use the method to convert to tensorRT, but the model can be used at creating-a-human-pose-estimation-application-with-nvidia-deepstream?

https://docs.nvidia.com/tlt/tlt-user-guide/text/deepstream_tlt_integration.html#installation-prerequisites

[kp2588417]

• I am able to generate the etlt model and tensorRT engine files by training the bpnet using COCO dataset.
• I want to run this model in the deepstream pipeline. Could you please let me know the process of running this in the deepstream pipeline. I mean is there any custom post-process script or osd script, I have to write so that I can be able to run?