Why channel-wisely compute feature map L2 distance?

[Yunzhi-b]

Hi, LPIPS helps me a lot in image translation task! There's one question that I could not figure it out by myself. For feature map distance, why the paper computes L2 distance channel-wisely and then averages spatially? Could we compute L2 distance spatially(flatten feature map for one channel, and compute L2 distance) and then average over channels?

Thanks!

[richzhang]

We average over the spatial dimensions, so that the metric is in roughly the same range of values, no matter the size of the image. The unweighted version is also equivalent to cosine distance in the channel direction.