ONNX model of RTMPose3D #3186
-
|
I am trying to achieve the same result as RTMPose3D's demo file using the converted ONNX model files to see if the files are suitable for a similar pipeline and if there was no issue during the conversion by writing a short pipeline. The detection part works fine, for the pose inference part, I receive a potentially promising output which is a list of tensors of following shapes: (1, 133, 576), (1, 133, 768), (1, 133, 576). 133 is promising because of keypoints but I'm not sure how to map this output into the visualization part to "move" the 3d skeleton. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
As far as I understand when the model is converted to onnx it doesn't convert the encoding decoding steps and it needs to be manually implemented. In this case if anyone else is faced with the same issue, find the class SimCC3DLabel's decode method and use that. |
Beta Was this translation helpful? Give feedback.
As far as I understand when the model is converted to onnx it doesn't convert the encoding decoding steps and it needs to be manually implemented. In this case if anyone else is faced with the same issue, find the class SimCC3DLabel's decode method and use that.