The frozen parameter in BN layer about temporal-segment-networks HOT 3 CLOSED

@GBJim
Basically, this can be interpreted as that RGB frames in videos are still a little different from static images on which the ImageNet models are trained. For example, consider motion blur and video compression which may affect the low-level visual features.
Using partial BN can help the model to account for these effects. This assumption is reflected in experimental results on split 1 of UCF101, which you can find in Table 1 of TSN paper.

from temporal-segment-networks.

Hi @kli-nlpr
It is a training technique called partial BN

You can find it on the original paper in the section3.2, Network Training part.

But I have a related question in here.
Based on the paper, Partial BN should be applied on Temporal CNN(flow), why is it also applied on Spatial CNN(RGB) as well?

from temporal-segment-networks.

Thank you for the detailed explanation! @yjxiong

from temporal-segment-networks.