Pilhyeon Lee (이필현)

I would like to suggest referring to Sec. 3.3. of the CMCS paper for the workflow of proposals ([link](https://openaccess.thecvf.com/content_CVPR_2019/papers/Liu_Completeness_Modeling_and_Context_Separation_for_Weakly_Supervised_Temporal_Action_CVPR_2019_paper.pdf)). The lambda and gamma were determined in a similar way with...

First of all, I apologize for the missing details about the experimental settings on ActivityNet. Since I extracted the features for ActivityNet by myself, they might be different from those...

The implementation of mmaction2 might be slightly different from ours. In ours, the temporal intervals are determined in an inclusive way, so we need to add 1 to compute areas...

Yes, I agree with you: it is more intuitive as the temporal slicing in python also works in that way. I will consider representing the intervals in that way in...

Hello, thanks for your interest! For the experiments on GTEA and BEOID, we utilized the features provided by SF-Net and you can find them [here](https://github.com/Flowerfan/SF-Net). On the other hand, we...

Hello, I provide some hyper-parameter settings for GTEA below. batch_size: 7 feature_fps: 15 search_freq: 1 budget: 200 lambdas: [1, 1, 1, 1, 1] Hope this would be helpful. Best regards.

Yes, we used the two-stream I3D networks pre-trained on Kinetics-400, as stated in the paper.

It is probably because we set the fps of all videos in THUMOS'14 to 25 before extracting the features.

Thanks for your suggestion! In fact, I have noticed some papers on fully-supervised temporal action localization that use such a label engineering technique. However, to my knowledge, existing weakly-supervised approaches...

As I have stated in the paper, we used the automatically generated point-level labels that are provided by [Moltisanti et al.](https://dimadamen.github.io/single_timestamps/paper/Action_Recognition_with_Single_Timestamp_Supervision__CVPR_.pdf) (CVPR'19). The point-level labels can be found on their...