actionformer_release Training and Evaluating On Our Own Dataset

Hi! thanks for your great work, look forward to your way to training and evaluating on our own dataset ?

Feb 23 '22 06:02 dulibubai

We are working on this and will probably update the repo after the ECCV deadline.

Feb 28 '22 17:02 happyharrycn

Hi, thanks for your work. When will you update the repo for training on our own dataset?

Mar 10 '22 06:03 csorg

We are still occupied with the ECCV supplementary materials. The plan is to update the repo (for customized training) in the upcoming weeks.

Mar 14 '22 05:03 happyharrycn

can you share feature extraction code on these public datasets?

Mar 23 '22 07:03 Darknesszlx

Hi, for THUMOS14 and ActivityNet dataset, we use the feature provided by CMCS. The CMCS uses pytorch-i3d-feature-extraction to extract the features. For TSP features, you may want to refer to TSP-official for details.

Mar 23 '22 07:03 tzzcl

Thanks for your reply!

Mar 23 '22 07:03 Darknesszlx

Sorry to bother you... I download the feature provided by CMCS, but I find that the feature extracted from THUMOS14 is different from yours. Where do I need to change to extract the same feature as yours? Besides, can you share i3d features extracted from ActivityNet dataset? Thank you very much! Looking forward to your reply~

May 23 '22 07:05 G0125

Hi, I think I directly use the feature of CMCS, If you downloaded the full part of CMCS, they provide multiple versions, I think I just choose the I3D version with ten-crop (I cannot remember the details). Also, the I3D features I'm used is also from CMCS repo

May 23 '22 07:05 tzzcl

Thank you! My problem has been solved!

May 23 '22 08:05 G0125

hi, I am trying to train your model on my own dataset and also applyed pytorch-i3d-feature-extraction to extract both RGB and flow features. I compared the features provied by CMCS and the results of I3D, then found they are in different shapes. Do do you have any idea that how rgb and flow features are fused? According to the I3D paper, the final prediction of an action detection task are fused by simply averages results of RGB I3D network and flow I3D network. As for feature extraction, are RGB features and flow features also proccsed in a similar way?

Jul 26 '22 02:07 fsi070614

For the action recognition task, you can directly use the average of RGB prob and Flow prob as the final action prediction. For the action detection task, the common practice is to concatenate the RGB feature and Flow feature into a final feature. For example, extracted features are 1024-d for both RGB and Flow modality. And the final feature dimension will be 2048.

Jul 26 '22 02:07 tzzcl

Thank you!

Jul 26 '22 02:07 fsi070614

Hi! I am looking forward to training and evaluating on our own dataset. It's Now July 26th. When will you update the repo for training on our own dataset? Thank you.

Jul 26 '22 08:07 CoderJackZhu

Hi, Thank you for your interest in our project!

Hopefully we will release that tutorial soon (no later than the ECCV conference).

Jul 27 '22 08:07 tzzcl

Hello, thank you for your interest in our project! You can directly send questions by E-mails as I can reply the email more effectively. My email is [email protected].

Aug 22 '22 10:08 tzzcl

hello, https://github.com/happyharrycn/actionformer_release/blob/main/libs/utils/train_utils.py#L426:~:text=results%20%3D%20postprocess_results(results%2C%20ext_score_file)请问这个后处理函数怎么用呢？ext_score_file指的是什么文件呢？

Sep 06 '22 07:09 dmmSJTU

For external scores, it means that we get the classification score for each video by external methods, i.e., a classification model. The postprocess_results is just fuse the external scores with the predictions from ActionFormer. I think you can simply discard this file when you train ActionFormer on your custom dataset.

Sep 06 '22 08:09 tzzcl

I got it. Thanks!

Sep 07 '22 09:09 dmmSJTU

+1 Looking forward to training and evaluating on our own dataset.

Oct 19 '22 07:10 lujiayou123

Is there an update on training and evaluating on our own dataset?

Jan 04 '23 08:01 Caspeerrr

Hi, thanks for sharing the code @happyharrycn. Could you let me know if you are still planning to share a recipe for training and evaluating on our own dataset?

Dec 09 '23 17:12 EmreOzkose

hello，I want to know how do you compute the GFLOPS? Can you share the code?

May 30 '24 09:05 caigouShaw

How to get a json file of a comment and how to training on our own dataset?

Jun 30 '24 10:06 010327Zy