add cpu mode

README.md

@@ -1,8 +1,10 @@
 # Learned-SIFT-Descriptor
 HW6 for 16-720
 # Prerequisites 
+* MATLAB
 * Anaconda
 * PyTorch: conda install pytorch torchvision -c soumith
 
 # To run the code
 run testMatch.m
+

computeLearnedSIFT.m

@@ -1,12 +1,9 @@
 function [locs,desc] = computeLearnedSIFT(im, GaussianPyramid, locsDoG, k, levels)
 [R,C,L] = size(GaussianPyramid);
-patchWidth = 9;
 locs = zeros(0,3);
-norm_patch_size = 64;
 patches=zeros(1,64,64);
 for i=1:size(locsDoG,1)
-    level = find(levels==locsDoG(i,3));
-    patch_half_size = floor(8*k^(1+levels(level)-levels(1)));
+
 %     bound = floor(sqrt(2)*patch_half_size
     bound = 32;
     if locsDoG(i,1)>=bound && locsDoG(i,2)>=bound && locsDoG(i,1)<=C-bound && locsDoG(i,2)<=R-bound

compute_desc.py

@@ -7,7 +7,7 @@
 import time
 import numpy as np
 def compute_desc():
-    net=SiameseNetwork().cuda()
+    net=SiameseNetwork()
 
     net.load_state_dict(torch.load('model.pt'))
     net.train(False)
@@ -20,10 +20,10 @@ def compute_desc():
         img=img.reshape(1,1,img.shape[0],img.shape[1])
         X=torch.from_numpy(img)
 
-        _,output=net(Variable(X.float()).cuda(),Variable(X.float()).cuda())
+        _,output=net(Variable(X.float()),Variable(X.float()))
 
         #print 'time:'+str(end-start)
-        mat[i]=output.data.cpu().numpy()
+        mat[i]=output.data.numpy()
     scipy.io.savemat('learned_desc.mat',mdict={'desc':mat})
     # print mat.tolist()
     return 0

compute_desc_cuda.py

@@ -0,0 +1,32 @@
+from net1 import SiameseNetwork
+from torchvision import transforms
+from PIL import Image
+import torch
+from torch.autograd import Variable
+import scipy.io
+import time
+import numpy as np
+def compute_desc():
+    net=SiameseNetwork().cuda()
+
+    net.load_state_dict(torch.load('model.pt'))
+    net.train(False)
+    # transform=transforms.ToTensor()
+    imgs=scipy.io.loadmat('patch.mat')['patches']
+    N=len(imgs)
+    mat=np.zeros([N-1,128])
+    for i in range(N-1):
+        img=imgs[i]
+        img=img.reshape(1,1,img.shape[0],img.shape[1])
+        X=torch.from_numpy(img)
+
+        _,output=net(Variable(X.float()).cuda(),Variable(X.float()).cuda())
+
+        #print 'time:'+str(end-start)
+        mat[i]=output.data.cpu().numpy()
+    scipy.io.savemat('learned_desc.mat',mdict={'desc':mat})
+    # print mat.tolist()
+    return 0
+if __name__ == "__main__":
+    compute_desc()
+

siftMatch.m

@@ -7,7 +7,7 @@
 %									into desc1 and the second column are indices into desc2
 
 if nargin<3
-    ratio = .8;
+    ratio = .6;
 end
 
 % compute the pairwise Hamming distance
@@ -16,8 +16,9 @@
 
 % suprress match between descriptors that are not distriminative.
 r = D(1,:)./D(2,:); 
-ix = ix((r < ratio & D(1,:)<0.5) | isnan(r)); 
-I2 = I(1,(r < ratio & D(1,:)<0.5) | isnan(r));
+
+ix = ix((r < ratio)& D(1,:)<2 | isnan(r)); 
+I2 = I(1,(r < ratio)& D(1,:)<2 | isnan(r));
 
 %output
 matches = [ix' I2'];

siftMatch.m~

@@ -7,7 +7,11 @@ function [matches] = siftMatch(desc1, desc2, ratio)
 %									into desc1 and the second column are indices into desc2
 
 if nargin<3
+<<<<<<< HEAD
     ratio = .;
+=======
+    ratio = .7;
+>>>>>>> 5dc9d5f515babe0140f91bc4ea2a6b668de3ceee
 end
 
 % compute the pairwise Hamming distance
@@ -16,8 +20,13 @@ ix = 1:size(desc1,1);
 
 % suprress match between descriptors that are not distriminative.
 r = D(1,:)./D(2,:); 
+<<<<<<< HEAD
 ix = ix(r < ratio | isnan(r)); 
 I2 = I(1,r < ratio | isnan(r));
+=======
+ix = ix((r < ratio)& D(1,:)<2 | isnan(r)); 
+I2 = I(1,(r < ratio)& D(1,:)<2| isnan(r));
+>>>>>>> 5dc9d5f515babe0140f91bc4ea2a6b668de3ceee
 
 %output
 matches = [ix' I2'];

testMatch.m

@@ -6,19 +6,17 @@
 % im1 = imread('../data/pf_scan_scaled.jpg');
 im1 = im2double(im1);
 if size(im1,3)==3
-    im1= rgb2gray(im1);
+    im1_gray= rgb2gray(im1);
 end
-[locs1, desc1] = learned_siftLite(im1);
+[locs1, desc1] = learned_siftLite(im1_gray);
 
 %im2 = imread('../data/chickenbroth_01.jpg');
 im2 = imrotate(im1,20);
+im2_gray=imrotate(im1_gray,20);
 % im2 = imread('../data/incline_R.png');
 % im2 = imread('../data/pf_stand.jpg');
 im2 = im2double(im2);
-if size(im2,3)==3
-    im2= rgb2gray(im2);
-end
-[locs2, desc2] = learned_siftLite(im2);
+[locs2, desc2] = learned_siftLite(im2_gray);
 
 [matches] = siftMatch(desc1, desc2);
-plotMatches(im1, im2, matches, locs1, locs2)
+img=getNewImg(im1, im2, matches, locs1, locs2);