search

doc-doc / NExT-QA

NExT-QA: Next Phase of Question-Answering to Explaining Temporal Actions (CVPR'21)
MIT License
114 stars 11 forks source link

Some confusions about extract my own pre-computed video features #22

Open dongfengxijian opened 1 year ago

dongfengxijian commented 1 year ago

Thank you for excellent work！I want to extract my own NExT-QA features. As the data preparation section of redeame, I download the the highlighted NExTVideo. When I unzip it，I don't know how to divide the raw video into train set, valid set and test set. The VidOR website also didn't provide the test set.

HU-xiaobai commented 1 year ago

hello, could you see all the videos are valid? I also download the raw videos but some of them are not valid, which means I could not open the video.

HU-xiaobai commented 1 year ago

@dongfengxijian

dongfengxijian commented 1 year ago

hello, could you see all the videos are valid? I also download the raw videos but some of them are not valid, which means I could not open the video.

I met the same problem. But when I extract frame with the scikit-video, it did‘t report error！Maybe the player is the reason,

HU-xiaobai commented 1 year ago

@dongfengxijian Hello, sorry to bother you again, could I ask if you could abstract the same appearance feature that the author provided? For example, for the validation set, I set the clip as 16 and I abstract the 8th or 9th(actually I compare all the frames in one clip) frame appearance vector of the first clip and compare the first frame vector of the provide appearance vector, both dimension are (2048,), but I find the appearance vector is different. how about yours?

HU-xiaobai commented 1 year ago

I now might could answer you question at the beginning. You could separate the video directly from the train.csv val.csv test.csv to make the video split

dongfengxijian commented 1 year ago

@dongfengxijian Hello, sorry to bother you again, could I ask if you could abstract the same appearance feature that the author provided? For example, for the validation set, I set the clip as 16 and I abstract the 8th or 9th(actually I compare all the frames in one clip) frame appearance vector of the first clip and compare the first frame vector of the provide appearance vector, both dimension are (2048,), but I find the appearance vector is different. how about yours?

Should the appearance vector be same? I don't quite grasp your question

dongfengxijian commented 1 year ago

I now might could answer you question at the beginning. You could separate the video directly from the train.csv val.csv test.csv to make the video split

Thank your for your guidance！

HU-xiaobai commented 1 year ago

@dongfengxijian , thanks for your answer! My question is: the appearance vector that I abstract is different from what the author provided, and how about you? Because we follow the same method, the appearance vector should be the same I think? because the weight of the resnext is fixed.

dongfengxijian commented 1 year ago

@dongfengxijian , thanks for your answer! My question is: the appearance vector that I abstract is different from what the author provided, and how about you? Because we follow the same method, the appearance vector should be the same I think? because the weight of the resnext is fixed.

I didn't use the original model, so I can't answer your question. I‘m sorry about that.

ghost commented 9 months ago

hello, could you see all the videos are valid? I also download the raw videos but some of them are not valid, which means I could not open the video.

I have the same issue