有关CTCN视频动作定位的问题

[happyCodingSusan]

谢谢作者的棒棒的成果,更谢谢作者分享您的代码.佩服佩服!

阅读了您的2018夺冠文章后,有两个问题请教,不知可否赐教? 您文章中说: BSN has a higher mAP than our C-TCN with the threshold of 0.95, however, it uses additional video classification results of [65] which are fused by multiple models, while our model only uses a two-stream model.

1. 您是否尝试过也用[65]的video classification results 替代您pipeline中two-stream model, mAP会提高吗?
2. 我阅读了[65] "Y. Zhao, et al., Cuhk & ethz & siat submission to activitynet challenge 2017", 发现这篇文章中介绍了好几个方法针对好几个问题.您说的video classification results究竟是哪种方法得到的?哪里可以下載its video classification results?

如果您能回覆, 就太感谢了!

[HUSTLX]

1、试过，不会，替换分类结果的话如果分类结果错误会全部错，CTCN自己生成的分类类别增加了部分召回。 2、这个是我们自己这边训了一些模型ensemble的，暂时不考虑release