add a new hash method 'PQ'

Method-SP/SP.m

@@ -0,0 +1,87 @@
+%function [R R2] = SP(X, bit, sparsity, n_iter)
+function [R R2] = SP(X, SPparam)
+%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
+% Solve:
+%   min |XR2'-B|^2 + BETA |XR2'-XR'|^2
+%   s.t. R2'R2=I, |R|0<= m
+% X: num * dim trianing data matrix. Note in our paper, X is dim * num.
+% R and R2: bit * dim projection matrix, R2 is orthogonal and R is sparse
+% B: num * bit binary codes
+%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
+
+bit = SPparam.nbits;
+sparsity = SPparam.sparsity;
+n_iter = SPparam.iter;
+
+
+% initialize with a random rotation
+dim = size(X, 2);
+R = randn(bit, dim);
+B = X*R';
+t = (B>0);
+B(t) = 1;
+B(~t) = -1;
+
+% pre compute pca matrix if bit < dim
+if(bit < dim)
+    [pc, ~] = eigs(cov(X),bit);
+    X_pc = X * pc;   
+end
+
+fprintf('iteration \n');
+beta = 1;
+for iter=0:n_iter
+
+    % fix B,R, update R2
+    Y = (B + beta*X*R')/(1+beta);
+    if(bit >= dim)
+        R2 = OrthogonalConstrainOpt(X,Y);
+    else
+        Rtmp = OrthogonalConstrainOpt(X_pc,Y);
+        R2 = Rtmp * pc';
+    end
+
+    % fix B,R2, update R
+    R = SparseConstrainOpt(R2, sparsity);
+
+    % fix R,R2, update B
+    B = X*R2';
+    t = (B>0);
+    B(t) = 1;
+    B(~t) = -1;
+end
+fprintf('\n');
+
+
+function R = OrthogonalConstrainOpt(X, Y)
+    %%% min |XR'-Y|^2, s.t. R'R=I
+    data_dim = size(X,2);
+    bit_num = size(Y,2);
+
+    [U Sigma V] = svd(X'*Y);
+    if(bit_num >= data_dim)
+        V = V(:,1:data_dim);
+    else
+        U = U(:,1:bit_num);    
+    end
+    R = V*U';
+end
+
+function R = SparseConstrainOpt(R2, sparsity)
+    n_total = numel(R2);
+    n_nonzero = ceil(n_total * (1-sparsity));
+    values = abs(R2(:));
+    values = sort(values, 'descend');
+    thresh = values(n_nonzero);
+    R=R2;
+    R(abs(R)<thresh) = 0;
+end
+
+end
+
+
+
+
+
+
+

Method-SP/SP.m~

@@ -0,0 +1,87 @@
+%function [R R2] = SP(X, bit, sparsity, n_iter)
+function [R R2] = SP(X, SPparam)
+%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
+% Solve:
+%   min |XR2'-B|^2 + BETA |XR2'-XR'|^2
+%   s.t. R2'R2=I, |R|0<= m
+% X: num * dim trianing data matrix. Note in our paper, X is dim * num.
+% R and R2: bit * dim projection matrix, R2 is orthogonal and R is sparse
+% B: num * bit binary codes
+%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
+
+bit = SPparam.nbits;
+sparsity = SPparam.sparsity;
+n_iter = SPparam.iter;
+
+
+% initialize with a random rotation
+dim = size(X, 2);
+R = randn(bit, dim);
+B = X*R';
+t = (B>0);
+B(t) = 1;
+B(~t) = -1;
+
+% pre compute pca matrix if bit < dim
+if(bit < dim)
+    [pc, ~] = eigs(cov(X),bit);
+    X_pc = X * pc;   
+end
+
+fprintf('iteration \n');
+beta = 1;
+for iter=0:n_iter
+
+    % fix B,R, update R2
+    Y = (B + beta*X*R')/(1+beta);
+    if(bit >= dim)
+        R2 = OrthogonalConstrainOpt(X,Y);
+    else
+        Rtmp = OrthogonalConstrainOpt(X_pc,Y);
+        R2 = Rtmp * pc';
+    end
+
+    % fix B,R2, update R
+    R = SparseConstrainOpt(R2, sparsity);
+
+    % fix R,R2, update B
+    B = X*R2';
+    t = (B>0);
+    B(t) = 1;
+    B(~t) = -1;
+end
+fprintf('\n');
+
+
+function R = OrthogonalConstrainOpt(X, Y)
+    %%% min |XR'-Y|^2, s.t. R'R=I
+    data_dim = size(X,2);
+    bit_num = size(Y,2);
+
+    [U Sigma V] = svd(X'*Y);
+    if(bit_num >= data_dim)
+        V = V(:,1:data_dim);
+    else
+        U = U(:,1:bit_num);    
+    end
+    R = V*U';
+end
+
+function R = SparseConstrainOpt(R2, sparsity)
+    n_total = numel(R2);
+    n_nonzero = ceil(n_total * (1-sparsity));
+    values = abs(R2(:));
+    values = sort(values, 'descend');
+    thresh = values(n_nonzero);
+    R=R2;
+    R(abs(R)<thresh) = 0;
+end
+
+end
+
+
+
+
+
+
+

demo.m

@@ -285,6 +285,23 @@
         [B_trn, ~] = compressBRE(H);
         [B_tst, ~] = compressBRE(H_query);
         clear db_data BREparam;
+
+     case 'SP' 
+         %Yan Xia,Kaiming He,Pushmeet Kohli,and Jian Sun.
+         %"Sparse Projections for High-Dimensional Binary Codes." In CVPR 2015.
+         addpath('./Method-SP/');
+         fprintf('......%s start...... \n\n', 'SP');
+         SPparam.nbits = param.nbits;
+         SPparam.sparsity = 0.9;
+         SPparam.iter = 50;
+         %train
+         R = SP(train_data,SPparam);
+         %coding
+         B_trn = (train_data*R' >=0);
+         B_tst = (test_data*R' >=0);
+         B_trn = compactbit(B_trn);
+         B_tst = compactbit(B_tst);
+         clear db_data SPparam;
 end
 
 % compute Hamming metric and compute recall precision

main_demo.m

@@ -54,7 +54,7 @@
     db_data = db_datalabel(:, 1:end);
 end
 
-hashmethods = {'LSH', 'ITQ'};
+hashmethods = {'SP', 'ITQ'};
 %hashmethods = {'Our Method', 'SELVE', 'CBE-opt', 'LSH', 'PCAH', 'SH', 'SKLSH', 'DSH', 'SpH'};
 %hashmethods = {'CBE-rand', 'CBE-opt', 'ITQ', 'LSH', 'PCAH', 'SH', 'SKLSH', 'PCA-RR', 'DSH', 'SpH'};
 nhmethods = length(hashmethods);