why res_manipulator returns enc_b + diff rather than enc_a + diff?
kkjh0723 opened this issue · 3 comments
kkjh0723 commented
In modules.py, I found that res_manipulator returns enc_b + diff rather than enc_a + diff (line 38).
According to the eq. 3 in the paper, isn't the later correct?
I might understand wrong...
12dmodel commented
It's possible we are using different definition of alpha in the code and in
the paper. When we return enc_b + diff, alpha = 0 means no change (and -1
for attenuated motion), while when we return enc_a + diff, this means alpha
= 1 means no change (and 0 for motion attenuated). In my experience there
is no real difference between the two, when alpha is large (say >10), the
difference doesn't seem to be noticeable.
I will double check these when I have time, thanks for pointing that out.
…On Fri, Nov 23, 2018 at 11:06 PM Jinhyung ***@***.***> wrote:
In modules.py, I found that res_manipulator returns enc_b + diff rather
than enc_a + diff (line 38).
According to the eq. 3 in the paper, isn't the later correct?
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#5>, or mute the thread
<https://github.com/notifications/unsubscribe-auth/ABtIV_kh252_5rfzei8Rkpepa40gn4Gxks5uyMWrgaJpZM4YxQ0Y>
.
kkjh0723 commented
Thank you. I understand.
ZhengPeng7 commented
To be very serious, it should only be enc_b + diff.
Suppose we have frame _ A/B/C/M in the given coco dataset, there is color perturbation between A and M. The perturbation was designed as a regularization for extracting motion better, instead of aiming to output an additional perturbation.
And if you check the amplification, enc_b + diff is also the one should be chosen.