How to use PCA_activitynet_v1-3.hdf5

[Darkbblue]

I'm trying to push the provided C3D features through a pretrained C3D model to possibly get some labels. I have noticed the features have been reduced from 4096 to 500 using PCA, so I guess I need to use this hdf5 to do an inversed PCA. I think I know the meaning of x_mean and U seems to be the eigenvectors. But no idea about what S is. It has 4096 dimensions and values in it seem very large.

[escorciav]

Thanks for your interest in our work. We noticed you're using an informally deprecated feature. Would it be possible for you to use more recent features instead, as they tend to yield better performance? I'll check my DAPs paper presented at ECCV2016 for code snippets regarding your question when time permits.

Sorry for any inconvenience.

[Darkbblue]

Thanks for your interest in our work. We noticed you're using an informally deprecated feature. Would it be possible for you to use more recent features instead, as they tend to yield better performance? I'll check my DAPs paper presented at ECCV2016 for code snippets regarding your question when time permits.

Sorry for any inconvenience.

Thanks for your reply. I have been reading the 2016 challenge website. Didn't notice it's deprecated... Should I follow this download page (http://activity-net.org/download.html) instead?

[escorciav]

What about if you use more recent features?

• We provided newer* features as part of the 2019 challenge,
• and 2022 challenge.

newer than C3D which it's from 2014. Actually, using C3D features is roughly equivalent to using AlexNet in 2023. Doesn't it sound outdated to you? :blush:

[Darkbblue]

What about if you use more recent features?

• We provided newer* features as part of the 2019 challenge,
• and 2022 challenge.

newer than C3D which it's from 2014. Actually, using C3D features is roughly equivalent to using AlexNet in 2023. Doesn't it sound outdated to you? 😊

Thanks. This helps a lot.