wangxiang1230/sstap

Code for our CVPR 2021 Paper "Self-Supervised Learning for Semi-Supervised Temporal Action Proposal".

sstap's Issues

thumos code

Thanks for your work! Could you share the code about thumos dataset.

Query regading the TSN feature

Hi , in the paper in methodology you have mentioned / cited a TSN feature which is pre-trained on Kinetics, however, in implementation details , you claim that for fair comparsion you have used same TSN feature as that of BMN/BSN . But BMN/BSN feature is fine-tuned on ActivityNet. WHich one is correct ?

Difference between train_semi_full and train_semi

Hi Author ! Congratulations on this exciting work getting accepted !

While going through the code I observed

1. SSTAP/main.py

   Line 145 in b8a55c9

   def train_BMN_Semi(data_loader, train_loader_unlabel, model, model_ema, optimizer, epoch, bm_mask):

2. SSTAP/main.py

   Line 359 in b8a55c9

   def train_BMN_Semi_Full(data_loader, model, model_ema, optimizer, epoch, bm_mask):

I am guessing what is the difference between this ? i observed that when unlabel percent is 0 you use train_semi_full , but how will it become semi-supervised if unlabel percent is 0% ? Is it experiment for ablation tables to show with 100% labels ?

if the mean-teacher parameters used on two datasets are the same？

Hello! I want to know if the mean-teacher parameters used on two datasets are the same.
eg.update_ema_variables(model, model_ema, 0.999, float(global_step / 20))
Thank you so much.

关于看您的源码中的一些疑问

作者您好，首先感谢您的论文给我带来了启发。在看您论文源码时，有些疑问想请教您。
我注意到您使用了csv_mean_100这个文件夹下的csv文件，请问这19228个csv文件是源视频数据集经过预训练后得到的特征文件嘛？
也就是说对应于论文图二中，经过 Encoding后得到的f1（即csv_mean_100这个文件夹下的csv文件分别对应了19228个视频样本的f1
特征）？

missing video features

请问如何使用i3d特征

你好，我看代码的数据处理，把特征都对齐为num_prop*400，然鹅i3d是2048维的特征，不知道应该如何处理。麻烦提供下使用i3d的pipeline ，谢谢

Hi, thanks for your contribution, when will the codes be avaliable

How can I print mAP values instead of AR@AN

Thank you so much for your excellent work. I am reproducing your model on my own dataset, I ask how can I print mAP values instead of AR@AN?

关于bm_mask()方法和add mask 操作的一些疑问

作者您好，关注您网络实现部分，发现对input_data_unlabel 数据生成的特征最后一步都进行了add mask 操作（teacher ，student 模型中都进行了）。请问add mask 操作的作用是什么？对应论文的哪一小节？ 为什么 input_data 数据没有实现add mask 操作？
此外，还有就是bm_mask()方法的作用是什么，bm_mask()和add mask 有什么区别?感谢答复

About the temporal action detection with generated proposals

Hi, thanks for sharing the code.
In the experiment, you have report the mAP performance of temporal action detection with the generated proposals.
Can you share the related code and procedure of adopting these propsals for action detection?
In several previous works, the details of this procedure are rarely introduced, usually described in the paper with a few words. As for student who is not familiar with the "detection by proposal classification" procedure, it is difficult to implement and calculate the performance. Hope you can share the details of it.
Thanks.

about i3d feature

Hello, before linear interpolation, is the feature of I3D extracted at the original frame rate or at the frame rate of 25FPS?
您好，请问在没有线性插值前，I3D的特征是以原始帧频提取的，还是以25fps的帧频提取的？

Can you explain what these variables mean: shift(x, n_segment, fold_div=8, inplace=False, channels_range=[1,2]).

shift(x, n_segment, fold_div=8, inplace=False, channels_range=[1,2]).
Now I want to use features extracted from I3D, but I do not konw what should I change, Can you tell me the meaning of these parameters? And if there are any other parameters that need to be adjusted, I would greatly appreciate it if you could let me know.

关于PRN

请问关于PRN的代码是否可以获得？

How can I print mAP values instead of AR@AN？

Thank you so much for your excellent work. I am reproducing your model on my own dataset, I ask how can I print mAP values instead of AR@AN?

关于数据扰动

您好，感谢您的两个工作。我是通过您PRN的工作了解到的数据扰动作为数据增强的操作，我能否问下在PRN中使用了那些方法来辅助数据扰动吗，因为在我使用其作为数据增强时，反而像您在SSTAP中关于TSM中操作的描述的一样,会导致性能下降。
谢谢您