Hello everyone, I need help to understand human pose estimation. I have followed the guide https://developer.nvidia.com/blog/creating-a-human-pose-estimation-application-with-deepstream-sdk/ but I cannot understand how to detect actions such as fall down.
How do I detect an action?
Hello, same problem -_-’ Any solution?Is there any tutorial that i can follow?
What’s the better approach? Action recognition or Human pose estimation
@pshin @dusty_nv
Regards
Rik
The sample in https://developer.nvidia.com/blog/creating-a-human-pose-estimation-application-with-deepstream-sdk/ just shows how to get the bodypose points with BodyPoseNet | NVIDIA NGC model. If you want to classify the bodypose points into some kind of “actions” or “movement”, you may need a classifier model which is similar to Pose Classification | NVIDIA NGC model. You need to train and deploy the new model by yourself.
Got it! Thank you @Fiona.Chen.
We analysed the datasets proposed by Nvidia as Kinetics but the class of human fall is not present (there are falling off bike and falling off chair, ref: PoseClassificationNet). We want a model that classifies the falling man on the ground (man down), is it correct to use the PoseClassificationNet? Is it appropriate to use other models? Can you suggest some solutions from which to start?
Whether the model can classify the “falling man” depends how you train the model. If you train the model with the ground-truth of falling man data, it can classify such “action” or “movement”. What we can provide is the model backbone and training tools. You need to do the training and figure out suitable solution according to your own requirement and resources.
Hi, we have used a post processing in Human Pose Estimation to find “man down” but the problem is finding a person with good accuracy. So we wonder if it’s better to use a top down technique. Is it possibile to use yolov7 pose estimation on deepstream? Or what do you suggest?
DeepStream is just a SDK. The accuracy is a model and algorithm problem. You can deploy any model with DeepStream in case the model is a ONNX, UFF, Caffe or TAO model.
For how to improve the model and algorithm accuracy, you may need to consult your experts for the model and algorithms. You may also need to retrain the model to handle your cases for all TAO pretrained models are not trained for you business cases. https://developer.nvidia.com/blog/training-custom-pretrained-models-using-tlt/
1 Like
Hi @Fiona.Chen , I am trying to develop the deepstream application for yolv7 pose estimation (in C++), but I cannot paraser the tensor from the NvDSInferTensorMeta. The prediction should contain this information: class_id, x, y , w, h , confidence and 51 keypoints (coordinates and confidence). I also saw the example in deepstream-infer-tensor-meta-test but it was not very helpful. Do you have any suggestions for me?
I m using Deepstream 6.0, Cuda 10.2 on Jetson NX.
1 Like
What does this mean? No NvDsInferTensorMeta available with your model or you don’t know how to parse the output layers?
Hi @Fiona.Chen thanks for the support, I’don’t know how parse the output layers in this case.
There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks
It depends on your model, please refer to the guys who provide the model to you.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.