search

google-deepmind / kinetics-i3d

Convolutional neural network model for video classification trained on the Kinetics dataset.
Apache License 2.0
1.75k stars 462 forks source link

Using I3D as a feature extractor for anomaly detection task #99

Closed hongsamvo closed 2 years ago

hongsamvo commented 4 years ago

I am using I3D as a feature extractor for anomaly detection task. Following the paper "Real world anomaly detection in surveillance videos" they use C3D which require an input chunk with 16 frames. As my understanding, I3D can run with every chunk > 8 frames. So I wonder that which number of frames per chunk I should use in this case? Thank you

joaoluiscarreira commented 4 years ago

Hi!

it should work with any number of frames (with minimal changes for very small number of frames). It should work better when using as many frames in evaluation as in training. So if the model was trained with 64 frames then it should work better evaluating on 64 frame clips.

Joao

On Tue, Mar 24, 2020 at 8:29 AM hongsamvo notifications@github.com wrote:

I am using I3D as a feature extractor for anomaly detection task. Following the paper "Real world anomaly detection in surveillance videos" they use C3D which require an input chunk with 16 frames. As my understanding, I3D can run with every chunk > 8 frames. So I wonder that which number of frames per chunk I should use in this case? Thank you

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/deepmind/kinetics-i3d/issues/99, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADXKU2SX2NKTVPSHE2LNFOLRJBVPVANCNFSM4LSO2LTA .

hongsamvo commented 4 years ago

Hi @joaoluiscarreira, Thank you so much for your responding