td-hm_hrnet-w32_coco doubt?

[chen-del]

Hello! I found that dev1.0 has changed a lot, but I have doubts about the pose output, what does its output mean, and its post-processing code is encapsulated. Where is the post-processing code for the actual reasoning? Please answer your doubts, thank you!

[Tau-J]

Hi @chen-del , thanks for using MMPose 1.0 ! According to your screenshot, the output feature is in shape [1, 17, 96, 72], which corresponds to 17 keypoint heatmaps, each with a size of 96x72. In MMPose 1.0, the post-processing operations are defined in the codec, which is a new module to encode/decode the targets. You can refer to this tutorial for details.

[chen-del]

Ok, thank you, but I saw that bbox is actually used in it, but I don’t know where to call it. If you want to use it, does poseonnx still need a pedestrian detection model to use bbox to return the coordinates of people?

[Tau-J]

@chen-del Sorry for late reply. Yes, for top-down algorithms, a pedestrain detector is needed.