Accuracy mismatch between document and test

[Tau-J]

Thanks for your error report and we appreciate it a lot. If you feel we have helped you, give us a STAR! :satisfied:

Checklist

1. I have searched related issues but cannot get the expected help.
2. The bug has not been fixed in the latest version.

Describe the bug

There's a significant mismatch between document and test result with Deeppose and RLE on MPII

	Head	Shoulder	Elbow	Wrist	Hip	Knee	Ankle	PCKh	[email protected]
RLE	Res50
Doc	95.80491	94.5822	86.89271	78.34604	87.46754	80.35467	73.45257	86.02394	26.31278
Master	87.82401091405184	94.49728260869566	87.23337152631714	79.61360259379691	87.38096624985977	79.70921574645614	67.7134269811466	84.88420504813948	26.278948799161096

Deeppose	Res50
Doc	93.55389	92.96875	82.73413	73.90828	85.65	75.41952	66.10711	82.53708	20.51002
Master	93.69031377899044	93.01970108695653	82.75121958021805	73.83997186045943	85.54614575872375	75.39928838568132	66.24886621884403	82.53187613843352	17.374446880038818

Reproduction

• following official tutorial, using res50 ckpt downloaded from document and run tool/test.py

ckpt:

https://download.openmmlab.com/mmpose/top_down/deeppose/deeppose_res50_mpii_256x256_rle-5f92a619_20220504.pth

https://download.openmmlab.com/mmpose/top_down/deeppose/deeppose_res50_mpii_256x256-c63cd0b6_20210203.pth

Environment

torch == 1.12.1+cu113 mmcv-full == 1.6.1

Bug fix

If you have already identified the reason, you can provide the information here. If you are willing to create a PR to fix it, please also leave a comment here and that would be much appreciated!