Why the test result on HRNet with DarkPose is the same as the result without DARK?

[superWSL]

Run with DarkPose's open source code. The test result on HRNet is the same as the result without DARK. There is no improvement. What is the reason? The configuration file and pre-training model are all used in HRNet.

[xizero00]

The reason is that there are some other modifications in our source code. If you use the original HRNet model provided by MSRA, these modifications will cancel the improvement from the DARK.

If you want to get the results for our DARK method, you should use the models we provided. If you want to evaluate the effect of the decoding part(our taylor method), you should copy the taylor function to the corresponding file in the original source code of HRNet modify this line and then use the model provided by MSRA.

[superWSL]

The reason is that there are some other modifications in our source code. If you use the original HRNet model provided by MSRA, these modifications will cancel the improvement from the DARK.

If you want to get the results for our DARK method, you should use the models we provided. If you want to evaluate the effect of the decoding part(our taylor method), you should copy the taylor function to the original source code of HRNet and then use the model provided by MSRA.

Thank you for your answer, I have received the pre-trained model you provided, and the test results are the same as those in the paper.

[superWSL]

Thank you, I have received the model you provided, and the test results are the same as those in the paper！

------------------ 原始邮件 ------------------ 发件人: "ilovepose/DarkPose" <notifications@github.com>; 发送时间: 2020年8月7日(星期五) 下午3:51 收件人: "ilovepose/DarkPose"<DarkPose@noreply.github.com>; 抄送: "FORGET"<1445934720@qq.com>;"Author"<author@noreply.github.com>; 主题: Re: [ilovepose/DarkPose] Why the test result on HRNet with DarkPose is the same as the result without DARK? (#9)

You should use the models we provided.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[seekFire]

@djangogo Hello! I wonder that what modification have you done to the original MSRA HRNet?

[xizero00]

@djangogo Hello! I wonder that what modification have you done to the original MSRA HRNet? You can use the diff tool to see the modifications

[seekFire]

@djangogo Thank you very much! Although using vimdiff, I can not find out any significant modification between your hrnet and official hrnet.

According to the opinion you mentioned:

If you use the original HRNet model provided by MSRA, these modifications will cancel the improvement from the DARK.

Can I understand that because your modified model doesn't perform as well as MSRA official model, official hrnet will have similar performance with modified hrnet + DARK ?