About the Synthetic Repetition

[Cecilia-yu412]

Hi @debidatta，i'm trying to figure out how the period length labels are generated when training RepNet, . As it described in the paper,"Then, we sample a clip C of random length P frames from V. This clip C is repeated K times (where K > 1) to simulate videos with repetitions."Is there anything worth noting when choosing the length of clip C and doing the inversion and splicing to ensure good training results? For example, the way of generating random numbers and the choice of the number of repetitions, I hope your answer.

[debidatta]

Yes you should ensure that the chosen segment gets repeated more than once. For example if you chose a segment of 32 frames or less it can repeat at least 2 times for the video length of 64. If you choose a longer segment it cannot repeat at least 2 times.

[Cecilia-yu412]

Yes you should ensure that the chosen segment gets repeated more than once. For example if you chose a segment of 32 frames or less it can repeat at least 2 times for the video length of 64. If you choose a longer segment it cannot repeat at least 2 times.

Thank you @debidatta, I got that. I have two other doubts, is it necessary to ensure that the number of duplicate clips of each synthetic video matches some distribution between 2~32 to prevent overfitting, or randint() has been sufficient? The second one is whether a large number of training sets are required for training? I want to synthesize the data on a training set of number about 800, roughly how many training samples should be generated， should i use a video for more than one time?

[debidatta]

is it necessary to ensure that the number of duplicate clips of each synthetic video matches some distribution between 2~32 to prevent overfitting, or randint() has been sufficient?

We used uniform sampling of period lengths. That is randint was sufficient.

The second one is whether a large number of training sets are required for training?

We were generating this data on the fly. For the final model we trained with batch size of 5 for 400k steps that is 2M synthetic repeating videos out of the 250k videos in Kinetics. That is about 8 synthetic videos per Kinetics video. You can possibly start with that or generate on the fly with more repeating videos per base video.