search

cvlab-yonsei / MNAD

An official implementation of "Learning Memory-guided Normality for Anomaly Detection" (CVPR 2020) in PyTorch.
333 stars 82 forks source link

why can't I achieve the accuracy in this article wtih my own trained model,while I can achieve it with yours #26

Open Pokeysx opened 3 years ago

fegler commented 3 years ago

Me too. I got 93.7, 84.7 on Ped2 and Avenue. But the paper results are 97.0, 88.5 each... Do you solve it?

Pokeysx commented 3 years ago

I only test his code on Avenue. As for pred task,I got 80.6 , and about reconstruction task，I got 77.9. The gap shocks me，so I give it up.

------------------ 原始邮件 ------------------ 发件人: "cvlab-yonsei/MNAD" @.>; 发送时间: 2021年3月30日(星期二) 中午12:47 @.>; @.**@.>; 主题: Re: [cvlab-yonsei/MNAD] why can't I achieve the accuracy in this article wtih my own trained model,while I can achieve it with yours (#26)

Me too. I got 93.7, 84.7 on Ped2 and Avenue. But the paper results are 97.0, 88.5 each... Do you solve it?

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

xiximin commented 3 years ago

Can you share the download link of shanghaitech dataset? I always fail to download

hyunjp commented 3 years ago

Hi, This link may help you. Link: https://github.com/cvlab-yonsei/MNAD/issues/1

Pokeysx commented 3 years ago

Thanks for your advice. when I use your pre-model, both the reconstruction task and the prediction task, the results are close to the results of the paper. However, When I used the model trained by myself, the results were not satisfactory. I tried to download data sets from the address you provide, reset the torch version==1.1.0 and increased the epochs,  the final AUC of the pred task was still 3-4 percent away from the paper.  If you have any more ideas, please do let me know. thanks again.

------------------ 原始邮件 ------------------ 发件人: "cvlab-yonsei/MNAD" @.>; 发送时间: 2021年4月1日(星期四) 晚上6:03 @.>; @.**@.>; 主题: Re: [cvlab-yonsei/MNAD] why can't I achieve the accuracy in this article wtih my own trained model,while I can achieve it with yours (#26)

Hi, I'm not sure what is the reason for the low performances, but this link may help you. https://github.com/cvlab-yonsei/MNAD/issues/1

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

zugexiaodui commented 2 years ago

The result of Ped2 AUC: 92.7425817630496 %

I also got a lower result on Ped2.

djy9809 commented 2 years ago

Sorry to disturb you. Have you solved this problem?

LiAoChUnLiNs commented 1 year ago

The label on the avenue may be incorrect. Many normal labels give abnormal.For example, the first one is 931 to 1100 frames of video.