search

martin-danelljan / Continuous-ConvOp

The Continuous Convolution Operator Tracker (C-COT).
GNU General Public License v3.0
196 stars 83 forks source link

regularization filter? #11

Open chenfsjz opened 5 years ago

chenfsjz commented 5 years ago

Hi, martin! In your CCOT project, dose the reg_filter in tracker.m filer denotes the matrix Wd in the equation (7) as below:

image

and another question is that why the size of Wd in your paper is (2Kd+2L+1)X(2Kd+1), because in your code, the size of reg_filter is 3x7.

can your take a little time to explain it?

martin-danelljan commented 5 years ago

Hi. Note that W is the convolution matrix corresponding to the convolution kernel w. The kernel w itself is 3x7 as you say. But W is the matrix corresponding to the linear operation of convolving the input with w. See our SRDCF paper for more details.

chenfsjz commented 5 years ago

Thank your! Ｍy understanding is that because circular convolution matrix C(hat(w)) in equation (6) of SRDCF paper is MNxMN corresponds to the Toeplitz matrix, the dimension of DFT(hat(w)) is MN, as the below equation shows,

image

I think that is ok. And when I focus on the regularization term in equation (7) in CCOT, image

you use the convolution kernel w to construct the convolution matrix Wd, and here Wd should also be a circular convolution matrix, if there has L non-zero number elements in it, I'm confused with the computation of its size: (2Kd+2L+1)X(2Kd+1).