search

MCG-NJU / RTD-Action

[ICCV 2021] Relaxed Transformer Decoders for Direct Action Proposal Generation
Apache License 2.0
86 stars 8 forks source link

a problem #21

Closed menghuaa closed 2 years ago

menghuaa commented 2 years ago

作者您好，anet代码在我的gpu上训练时间太久，整个流程大概需要5天，这对我来说试错成本太高了。 想问下您，activitynet1.2是activitynet1.3的子类，如果我用RTD在activitynet1.2上训练，从activitynet1.3的视频特征中提取出属于activitynet1.2的视频的特征，这样是否是可以的呢？因为目前tsn的特征，只有activitynet1.3的，没有activitynet1.2的。然后我也将activitynet1.2与activitynet1.3的注释文件进行了对比，结果显示activitynet1.2的视频在activitynet1.3都存在，且视频持续时间都是一样的，每个视频annotations的数量都是一致的，仅有三个视频activitynet1.3的annotations的数量增加了一个。

JackyTown commented 2 years ago

I think ActivityNet-1.2 is a subset of ActivityNet-1.3, thus it is okay.

JackyTown commented 2 years ago

ActivityNet-v1.2 has not been mentioned for a long time.