some problems about technique: udp and darkpose

[hatecodes]

When I train HRNet with udp and darkpose, the results are far lower than the original HRNet：

original HRNet:
"AP": 0.74573, "AP .5": 0.80809, "AP .75": 0.7267,
with udp:
"AP": 0.67795, "AP .5": 0.7729, "AP .75": 0.66005,
with darkpose:
AP: 0.7357, AP .5: 0.7901, AP .75: 0.7499

im curious about that, plz help, thx a lot!

[Tau-J]

Hi, thanks for using MMPose. Could you please provide more information about your training settings, config, datasets, env etc.?

[hatecodes]

Hi, thanks for using MMPose. Could you please provide more information about your training settings, config, datasets, env etc.?

HI TauJ, im currently using the default training settings. and my own made datasets. env: windows+python3.7+mmpose0.24.0 CPU is AMD 5900X and GPU is RTX3090

configs are:
mmpose-master\configs\body\2d_kpt_sview_rgb_img\topdown_heatmap\coco\hrnet_w48_coco_384x288.py
mmpose-master\configs\body\2d_kpt_sview_rgb_img\topdown_heatmap\coco\hrnet_w48_coco_384x288_dark.py
mmpose-master\configs\body\2d_kpt_sview_rgb_img\topdown_heatmap\coco\hrnet_w48_coco_384x288_udp.py

[Tau-J]

Sorry, for self-made datasets, I can't give any suggestions.