Human Pose Estimation Application improvement required

Please provide complete information as applicable to your setup.

• Hardware Platform (GPU)
• DeepStream Version 5.0
• TensorRT Version 7.2.1-1
• NVIDIA GPU Driver Version (valid for GPU only) 450.102.04
• Issue Type (questions)

Human Pose Estimation Application improvement required

I tried to run the human pose estimation from this blog post(Creating a Human Pose Estimation Application with NVIDIA DeepStream | NVIDIA Developer Blog). It is also pinned in this forum as well(Pose Estimation with DeepStream).

I get good fps in jetson and gpu laptop but the drawing accuracy is very poor when there are multiple persons in the image. example:


(ignore the bounding box) In this picture we can see that it detects key points but links them very poorly. I tried to change the link_threshold and threshold and check if the situation improved or not. But unfortunately these did not improve the matter. I have tried both resent18 and densenet121 pre-trained weights.

How can I improve these linking problem in trt pose models?


Hey customer, have you modified the original GitHub - NVIDIA-AI-IOT/deepstream_pose_estimation: This is a sample DeepStream application to demonstrate a human pose estimation pipeline. , could you share your video and other things if have, I will repro and debug locally.

I have the same issue as stated here: Deepstream_pose_estimation
So I used normal onnx conversion way to convert to onnx by myself. Then this type of drawing is always generated.