chrisallenming / cross_category_video_highlight Goto Github PK

I tried to download these two data sets through the method you provided, but they couldn't be downloaded for various reasons. Can you provide these two data sets, for example, through a network disk?

Hello. Thanks for the great work.

While we were checking the highlight video dataset, we have found that some videos are not in accordance with their video labels.

For example, 'dog/iyYiqa0QZXM/' video in youtube is the video illustrating how to draw monkey.

The video below is also about drawing the zebra but included in the dog class.
'http://www.youtube.com/watch?v=7H_qqQvPUzY'

Could you provide how the authors handled these?
Or weren't there any considerations for these files?

Thank you.

hi thanks for share your research on videohight domain，
I train the SL_module on with src_category=surfing and tgt_category=surfing and got the results map=0.67,however in the paper The same configuration is map=0.76, the tensorboard results is shown below：

Could you tell me if there is something wrong with my training？

There is a pretained C3D model and no trained DL-VHD model. Is it possible to open the DL-VHD model so we can download and test it

Dear author:
Thanks for the DA on VHD work. I wonder have you tested the predicted HL curves on real scenario video, such as the videos captured by mobile phone on daily scene. Does it work also good and reasonable? thank you for giving some feedback about this.

Thanks for your work! I have something not sure about experiments of the activitynet dataset.

1. In your paper, you said the label is tIoU, so is that a float data or an integer? I suppose tIoU to be a float data, but in your code, it seems to be an integer (in class ActivityNet_Set in activity_net_set.py). What's more, if the label is an integer, why not use "np.round" instead of "np.array"?
2. Is the 3k+ video data in activitynet you used is randomly selected from the whole dataset? Will more videos lead to a better result?

您好 ，我在通过你的模型进行网络训练时，有一个疑惑就是关于红色框中的输入，cls src_cls tgt_cls分别代表什么呢 我看了一下源码，他们都是['dog', 'gymnastics', 'parkour', 'skating', 'skiing', 'surfing'] 列表中的其中一个，所以我没有明白Source-only Target-oracle. DA baselines.的区别是啥，或者说怎么得到的，希望您能告诉我 在运行Source-only Target-oracle. DA baselines.的Python命令时，--src_category --tgt_category 这两个参数应该分别输入什么呢？ 我觉得不可能是将 ['dog', 'gymnastics', 'parkour', 'skating', 'skiing', 'surfing']进行两两组合进行训练，这样的话会产生很多模型。

Hello, thanks for sharing the code.

After I tested the SL_module on the "skating" category, I could only get an mAP around 30%. There were also some gaps of mAP on other categories between my evaluations and what has been reported in the paper, but they were not as ridiculous as for "skating".

So I guess there might be some discrepancy between the set of videos we used for training/testing, since I guess the videos are constantly disappearing from youtube. So could you share the train/test video ids??

Or do you have any other clues about this problem on "skating"?

Thanks.

Hi @ChrisAllenMing, thanks for your great work on video highlight detection! As far as I know, ~190 videos are missing in the YouTube Highlights dataset currently. May I ask that how did you handle this problem? Thanks!

Yours paper says there are 2520 videos for training and 1260 videos for test, while there are only 3317 videos in ActivityNet_vids.txt in total. What is the reason for this? If I train model with these, will it affect the results?