There were two bugs in the sample code, which leads to broken results:
1. In MGutils, line 444, only x channel is read, now it's fixed to three channels.
2. In MGutils, line 324, the parameters for the attention operation were set as -15.5 and 6.0, which should be as the IJCV paper reported: -1.0/3.0.
3. Feature extraction function: Extract_moving_pose_Feature is replaced with Extract_feature_Realtime to be consistent with the defined feature num in line 68, (though Extract_moving_pose_Feature gives better results by our experience).
Posted by: mikecheninoulu @ May 9, 2023, 11:16 p.m.