Open niupihan opened 5 years ago
那么P转置乘P等于P这个等式和BACF中不一致,并不会有这样的结果啊?除非P等于单位矩阵
P是一个diagonal binary matrix,维度是T*T,T是feature map一个channel里的pixel个数,例如feature map是50*50*256,T=50*50=2500,P就是一个2500*2500的diagonal binary matrix,是满足(P^T)P=P的。
谢谢
请问代码什么时候能分享出来啊
请问ADMM_base里面对h的求解分母怎么少了w呢
adaptive spatial regularization 只在定位的时候用。
in ADMM_solve_w(params,use_sz,model_w,h_f) ..... w = bsxfun(@rdivide,(q-m),(1+(params.admm_lambda1/mu)*Hh));
is the same as equation (13) ????
w has different types of solution, the paper gives the analytical solution of the w sub-problem, the code is to use ADMM solution, this is because we found that using ADMM will get better results during the experiment, we also explained in the paper --"In practice, we utilize an additional ADMM solver to obtain the weight w∗ for better convergence".
If you want to understand the ADMM solution of w, you can refer to the following:
@Daikenan ,Can you tell me which paper the "ADMM for sloving w" is from?Thank you.
This part of the solution is written by me and corresponds with code.
@Daikenan Hello,I have read the code,I have some questions,why do you execute the ADMM algorithm when the frame is even number,why not uses ADMM algorithm for each frame.I remember that the paper ECO_HC adopts the similar step,I wonder why you do this.Thank you.
@Daikenan Hello,I have read the code,I have some questions,why do you execute the ADMM algorithm when the frame is even number,why not uses ADMM algorithm for each frame.I remember that the paper ECO_HC adopts the similar step,I wonder why you do this.Thank you.
It has two effects, one is to improve the speed, making sure to achieve real time speed. The other is to reduce over-fitting.
@Daikenan Thanks for your reply.First question: according to the code,the variables of model_w,g_hcf,g_f would not update when the frame is odd number. Second question: Have you ever tested that it works best when the number of frames is even number,for example,Have you ever done that it works when the variable "update_interval" equals 3 or 4?if you did it,Can you tell me what the result is?Thanks a lot.
wr是reference spatial regularization weights,论文4.2 model update里是有说明的,初始化是负高斯,用线性加权w的方式进行更新,程序里最后lr是1,即上一帧的w。