pooling strategy about videval

[vztu]

This is a great question. My procedure in this pooling paper is (take BRSIQUE as example):

1. First extract 36 features per frame for a video, yielding a total of 300 x 36 frame features, supposing each video has 300 frames.
2. Average pool frame features for each video, resulting in 1 x 36 features for each video.
3. Do this on the training set (eg, 500 videos), you get a 500 x 36 feature matrix, and then train the SVR on MOSs. Now you have a quality predictor for any 36-dim BRISQUE feature (Even though trained on videos)
4. Inference phase on the test set (eg, 100 videos): for a given video with 300 frames, extract 300 x 36 BRISQUE features, then apply the trained SVR model, you get 300 x 1 quality predictions. Now you can apply your favorite (training-free) temporal pooling methods to reduce these time-series scores!

Another method is to assign the video score to each frame and then train on 300x more samples. I quickly tried this one and found that it didn't give better results, but training is much slower since you get 300x more training samples. These two methods both have flaws however. I think a better way is to train the implicit temporal pooling in an end-to-end manner, such as VSFA, or CoINVQ

[vztu]

For your reference, we have open-sourced several pooling methods here: https://github.com/vztu/Temporal_Pooling

[kobe233333]

For your reference, we have open-sourced several pooling methods here: https://github.com/vztu/Temporal_Pooling

Appreciate a lot! I will have a try.

[kobe233333]

Another question is whether SVM can be replaced by other regression methods, such as MLP, when performing regression after obtaining features？

[vztu]

Yes you can use your favorite regressor. My experience is on relatively small dataset (<5k), SVM usually performs well

…

On Fri, Oct 8, 2021 at 2:01 AM kobe233333 ***@***.***> wrote: Another question is whether SVM can be replaced by other regression methods, such as MLP, when performing regression after obtaining features？ — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#11 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AKKGPFQ5CHX3FKMVCEVBJOLUF2JN5ANCNFSM5D7S6KMA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub> .

[kobe233333]

Yes you can use your favorite regressor. My experience is on relatively small dataset (<5k), SVM usually performs well
…
On Fri, Oct 8, 2021 at 2:01 AM kobe233333 @.***> wrote: Another question is whether SVM can be replaced by other regression methods, such as MLP, when performing regression after obtaining features？ — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#11 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AKKGPFQ5CHX3FKMVCEVBJOLUF2JN5ANCNFSM5D7S6KMA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub .

Thanks！