the training set of multithumos?

[Wyz2927]

hello, how about the training set of multithumos? Is it the same as charades? I have trained my own model using the following training set, but only get 40.2136 f-map. Do you konw what is the problem?

[dairui01]

Thanks for the interest in the MS-TCT.
From the provided parameters, it may need to change the batch size and initial learning rate.
You can try batch_size = 8 and lr = 0.0005. As MultiTHUMOS has a difference in terms of video length compared to Charades, so you may need to change the number of batch sizes. Also, the I3D features that are not pre-trained on the MultiTHUMOS training set may also hamper the performance.

[Wyz2927]

Thanks for the interest in the MS-TCT. From the provided parameters, it may need to change the batch size and initial learning rate. You can try batch_size = 8 and lr = 0.0005. As MultiTHUMOS has a difference in terms of video length compared to Charades, so you may need to change the number of batch sizes. Also, the I3D features that are not pre-trained on the MultiTHUMOS training set may also hamper the performance.

Thank you so much for your quick reply. I have trained my model with your suggested setting and I got 43.8647% full-val-fmap at the 53rd epoch, even higher than you reported in the paper.

And I have another question. I found that you also report the map score of MLAD on multithumos in your paper. I have tried to reproduce this results but only get around 37.5 map. Could you please share the experimental setup if it is convenient?
I use the following settings.

[dairui01]

Thanks for testing the parameters. As I have refined a version of the code after the CVPR submission, thus it may perform better than the one reported in the paper.

Regarding the MLAD on MultiTHUMOS, I am also unable to reproduce the result reported in that paper. So in the table of "comparison with SOTA", I just reuse their paper's number directly. Maybe you can raise this issue in MLAD's repo.

[wzy246532]

Dear Wyz2927,

I have seen the results of your training on the MultiTHUMOS dataset using the MS-TCT model. I would appreciate it if you could share the feature data extracted by MultiTHUMOS on I3D for MS-TCT input.

Best wishes!

[Wyz2927]

Dear Wyz2927,

I have seen the results of your training on the MultiTHUMOS dataset using the MS-TCT model. I would appreciate it if you could share the feature data extracted by MultiTHUMOS on I3D for MS-TCT input.

Best wishes!

Another paper MLAD has provided the features for MultiTHUMOS. you can find it in https://github.com/ptirupat/MLAD.

[wzy246532]

------------------ 原始邮件 ------------------ 发件人: "dairui01/MS-TCT" ***@***.***>; 发送时间: 2022年12月1日(星期四) 下午5:03 ***@***.***>; ***@***.******@***.***>; 主题: Re: [dairui01/MS-TCT] the training set of multithumos? (Issue #2) Dear Wyz2927, I have seen the results of your training on the MultiTHUMOS dataset using the MS-TCT model. I would appreciate it if you could share the feature data extracted by MultiTHUMOS on I3D for MS-TCT input. Best wishes! Another paper MLAD has provided the features for MultiTHUMOS. you can find it in https://github.com/ptirupat/MLAD. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented. Thank you very much for your quick reply. I have downloaded the corresponding rgb features, which are rgb_test.h5 and rgb_val.h5 respectively. But I ran into a new problem, they are different from charades' features, I don't know how to deal with them? Are they treated in the same way as charades features? Best wishes!

[shiyi-z]

hi the link of the features for MultiTHUMOS provided by MLAD was invalid, I wonder that do you have a new link or can you share the profile of the data feature ,i would appreciate it !

[1220894232]

Hi there,
I wanted to ask you about the MultiTHUMOS dataset and its features that you've shared on your GitHub repository at https://github.com/ptirupat/MLAD. I've been trying to access the features through the provided link, but it appears to be broken. I was wondering if there is a new link available or another way to obtain these features?
I appreciate any help you can provide. Thank you in advance.
Best regards