About action detection with proposals
Rheelt opened this issue · comments
Hello, thank you for your contribution.
In the paper, On ActivityNet-1.3, we use top-1 video-level class generated by classification model [44] for all proposals in a video. ([44] means UTS at ActivityNet 2016). But I can't find the source code.
So my question is how do you actually do action detection with proposals?
Thank you very much!
Hi @coder-LeeTao , we obtain [44]'s result files by e-mail in last year. Recently, I found that these result files may have some problem. Thus we are replacing it with classification results of another solution (CUHK at ActivityNet 2017), and will update our arxiv paper in several days.
If you also need their classification results for action detection, you can send a email to [CUHK at ActivityNet 2017].
Please feel free to reopen this issue if you have furthur problem.
Thank you very much @wzmsltw ! I have another question. In the paper, On THUMOS14, we use top-2 video-level classes generated by UntrimmedNet [43] for proposals generated by BSN and other methods, where multiplication of confidence score and class score is used for retrieving detections.
My understanding is using in NMS retrieval. The proposal is then assigned to the action category with the highest score in UNet.
Actually, top-2 means we combine two classes with a proposal separately. So the first proposal result with class_1 has score p_f = p_{conf} * p^s * p^e * p_{class_1}, and the second proposal result with class_2 has score p_f = p_{conf} * p^s * p^e * p_{class_2}. And we actually use NMS on proposals first, then combine proposals with action classes.
Thank you very much@wzmsltw ! So in the final result, is a proposal used twice(with different action class label)?
Yes