Where is the "Temporal clipping" part ?

[DK-Jang]

Hello, I am very interested in you paper :) While looking at your code, I can't find the "temporal clipping" part (paper page 7) for data augmentation in your code. I want to know where is the implementation about " in every iteration we randomly select the temporal length from the set T ∈ {64, 56, 48, 40}"

Thank you.

[ChrisWu1997]

Thank you for looking into our paper:)

Sorry I forget to put the temporal clipping part in the code. You can try to add the following lines in the forward function of agent.py here:

temp_lens = [64, 56, 48, 40]
tlen = temp_lens[random.randint(0, 3)]
start = random.randint(0, 64 - tlen)
inputs = [x[:, :, start:start+tlen] for x in inputs]
targets = [x[:, :, start:start+tlen] for x in targets]

In general, this temporal clipping augmentation is not crucial.

[rozentill]

Hi, since you only trained on such length sequences, is it possible to extract features from a very long input sequence? For example a one minute video which would have 1000+ frames. Thanks.

[ChrisWu1997]

It's OK to use long sequence input, as it's a fully convolutional network. And notice that we use ReflectedPadding for all conv layers, which we found beneficial to apply the system on long sequence. If using ZeroPadding, indeed there will be artifacts when using long sequence.