retrain

SiameseNetworkDataset.py

@@ -2,25 +2,52 @@
 import torch
 import numpy as np
 from PIL import Image
+
+import random
 class SiameseNetworkDataset(Dataset):
-    def __init__(self, root_dir, label, transform=None):
-        self.root_dir = root_dir
-        self.label = label
+    def __init__(self, imageFolderDataset, transform=None):
+        # self.root_dir = root_dir
+        # self.label = label
+        # self.transform=transform
+        self.imageFolderDataset=imageFolderDataset
         self.transform=transform
     def __getitem__(self, index):
 
-        img0 = Image.open(self.root_dir + '/pair1/' + str(index) + '.bmp')
-        img1 = Image.open(self.root_dir + '/pair2/' + str(index) + '.bmp')
+        # img0 = Image.open(self.root_dir + '/pair1/' + str(index) + '.bmp')
+        # img1 = Image.open(self.root_dir + '/pair2/' + str(index) + '.bmp')
+        #
+        # img0=img0.convert("L")
+        # img1=img1.convert("L")
+        #
+        # if self.transform is not None:
+        #     img0=self.transform(img0)
+        #     img1=self.transform(img1)
+        #
+        # return img0,img1, torch.from_numpy(np.array([int(self.label[index])],dtype=np.float32))
+
+        img0_tuple=random.choice(self.imageFolderDataset.imgs)
+        should_get_same_patch=random.randint(0,1)
+        image0=Image.open(img0_tuple[0])
+        if should_get_same_patch:
+            image1=image0
+        else:
+            img1_tuple = random.choice(self.imageFolderDataset.imgs)
+            while img1_tuple[0]!=img0_tuple[0]:
+                break
+            image1=Image.open(img1_tuple[0])
+
+        x=random.randint(0,359)
+        image1=image1.rotate(x)
 
-        img0=img0.convert("L")
-        img1=img1.convert("L")
+        image0 = image0.convert('L')
+        image1 = image1.convert('L')
 
         if self.transform is not None:
-            img0=self.transform(img0)
-            img1=self.transform(img1)
+            image0=self.transform(image0)
+            image1=self.transform(image1)
+        return image0, image1, torch.from_numpy(np.array([should_get_same_patch],dtype=np.float32))
 
-        return img0,img1, torch.from_numpy(np.array([int(self.label[index])],dtype=np.float32))
     def __len__(self):
-        return len(self.label)
+        return len(self.imageFolderDataset.imgs)
 
 

computeLearnedSIFT.m

@@ -13,19 +13,21 @@
         % compute 128-D discriptor
 
         patch = im(locsDoG(i,2)+1-bound:locsDoG(i,2)+bound,locsDoG(i,1)+1-bound:locsDoG(i,1)+bound);       
+        %patch = GaussianPyramid(locsDoG(i,2)+1-bound:locsDoG(i,2)+bound,locsDoG(i,1)+1-bound:locsDoG(i,1)+bound,level);       
+
         % compute major gradient
-        tmp_patch = GaussianPyramid(locsDoG(i,2)+1-patch_half_size:locsDoG(i,2)+patch_half_size,locsDoG(i,1)+1-patch_half_size:locsDoG(i,1)+patch_half_size,level);
-        tmp_patch = imresize(tmp_patch, [norm_patch_size norm_patch_size]);
-        g_histogram = computeGradient(tmp_patch);
-        [~,major_g_index] = max(g_histogram);
+        %tmp_patch = GaussianPyramid(locsDoG(i,2)+1-patch_half_size:locsDoG(i,2)+patch_half_size,locsDoG(i,1)+1-patch_half_size:locsDoG(i,1)+patch_half_size,level);
+        %tmp_patch = imresize(tmp_patch, [norm_patch_size norm_patch_size]);
+        %g_histogram = computeGradient(tmp_patch);
+        %[~,major_g_index] = max(g_histogram);
         % rotate patch
-        patch = imrotate(patch,-45*(major_g_index-1));
-        [h,w] = size(patch);
-        h = floor(h/2);
-        w = floor(w/2);
-        patch = patch(h+1-patch_half_size:h+patch_half_size,w+1-patch_half_size:w+patch_half_size);
-        patch = imresize(patch, [norm_patch_size norm_patch_size]);
-        patches(end,:,:)=patch;
+        %patch = imrotate(patch,-45*(major_g_index-1));
+        %[h,w] = size(patch);
+        %h = floor(h/2);
+        %w = floor(w/2);
+        %patch = patch(h+1-patch_half_size:h+patch_half_size,w+1-patch_half_size:w+patch_half_size);
+        %patch = imresize(patch, [norm_patch_size norm_patch_size]);
+        patches(end,:,:)=imresize(patch,[64,64]);
         patches=[patches;zeros(1,64,64)];
     end
 end

compute_desc.py

@@ -10,15 +10,20 @@ def compute_desc():
     net=SiameseNetwork()
 
     net.load_state_dict(torch.load('model.pt'))
-    transform=transforms.ToTensor()
+    net.train(False)
+    # transform=transforms.ToTensor()
     imgs=scipy.io.loadmat('patch.mat')['patches']
-    imgs=imgs.reshape(imgs.shape[0],1,imgs.shape[1],imgs.shape[2])
-    imgs=np.delete(imgs,[imgs.shape[0]-1],axis=0)
-    X=torch.from_numpy(imgs)
-    _,output=net(Variable(X.float()),Variable(X.float()))
+    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)
 
-    #print 'time:'+str(end-start)
-    mat=output.data.numpy()
+        _,output=net(Variable(X.float()),Variable(X.float()))
+
+        #print 'time:'+str(end-start)
+        mat[i]=output.data.numpy()
     scipy.io.savemat('learned_desc.mat',mdict={'desc':mat})
     # print mat.tolist()
     return 0

cropdata.m

@@ -0,0 +1,20 @@
+img=imread('../data/book.jpg');
+
+img=imrotate(img,-90);
+
+img=imresize(img,0.1);
+
+img=rgb2gray(img);
+
+[h,w]=size(img);
+
+patch=zeros(64,64);
+counter=0;
+for x=1:10:w-63
+    for y=1:10:h-63
+        patch=img(y:y+64-1,x:x+64-1);
+        imwrite(patch,sprintf('book/%d.jpg',counter));
+        counter=counter+1;
+    end
+end
+

cropdata.m~

@@ -0,0 +1,17 @@
+img=imread('../data/book.jpg');
+
+img=imrotate(img,-90);
+
+img=imresize(img,0.1);
+
+img=rgb2gray(img);
+
+[h,w]=size(img);
+
+patch=zeros(64,64);
+for x=1:2:w-63
+    for y=1:2:h-63
+        patch=img(y:y+64-1,x:x+64-1);
+    end
+end
+