add new stuff

compute_desc.py

@@ -7,7 +7,7 @@
 import time
 import numpy as np
 def compute_desc():
-    net=SiameseNetwork()
+    net=SiameseNetwork().cuda()
 
     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()),Variable(X.float()))
+        _,output=net(Variable(X.float()).cuda(),Variable(X.float()).cuda())
 
         #print 'time:'+str(end-start)
-        mat[i]=output.data.numpy()
+        mat[i]=output.data.cpu().numpy()
     scipy.io.savemat('learned_desc.mat',mdict={'desc':mat})
     # print mat.tolist()
     return 0

getNewImg.m

@@ -0,0 +1,29 @@
+function newImg = getNewImg(myim1, myim2, matches, locs1, locs2)
+% INPUT
+% im1, im2 - grayscale images, scaled into [0 1]
+% matches -  the list of matches returned by briefMatch()
+% locs1, locs 2 - the locations of keypoints returne by briefLite() . 
+%                 m1 x 3 and m2 x3 respectively. each row corresponds to the 
+%                 space-scale position: [x_coordinate,y_coordiante,scale]
+
+
+    im1= im2double(myim1);
+    im2= im2double(myim2);
+
+    width = size(im1,2) + size(im2,2);
+    height = max(size(im1,1), size(im2,1));
+    nchannel = size(im1,3); 
+    img = zeros(height, width,nchannel);
+%     size(img)
+    img(1:size(im1,1),1:size(im1,2),:) = im1;
+    img(1:size(im2,1),size(im1,2)+1:size(im1,2) + size(im2,2),:) = im2; 
+
+    position = zeros(size(matches,1),4);
+
+    for i = 1:size(matches,1)
+        p1 = locs1(matches(i,1),:);
+        p2 = locs2(matches(i,2),:); 
+        position(i,:) = [p1(1) p1(2) size(im1,2)+p2(1) p2(2)];
+    end
+    newImg = insertShape(img,'Line',position,'Color','r','LineWidth',1);
+end

new_testMatch.m

@@ -0,0 +1,44 @@
+myim1 = imread('../data/book.jpg');
+myim1 = imrotate(myim1,-90);
+myim1 = imresize(myim1,0.1);
+im1 = im2double(myim1);
+if size(im1,3)==3
+    im1= rgb2gray(im1);
+end
+[locs1, desc1] = learned_siftLite(im1);
+% 
+% im2 = imrotate(im1,45);
+% [locs2, desc2] = briefLite(im2);
+% [matches] = briefMatch(desc1, desc2);
+% plotMatches(im1, im2, matches, locs1, locs2);
+
+video_file='../data/IMG_0918.mp4';
+video=VideoReader(video_file);
+frame_number=floor(video.Duration * video.FrameRate)-1;
+
+myvideo = zeros(384,461,3,frame_number);
+for i=1:frame_number
+    myim2 = read(video,i);
+    myim2 = imresize(myim2,0.2);
+    im2 = im2double(myim2);
+    if size(im2,3)==3
+        im2= rgb2gray(im2);
+    end
+    [locs2, desc2] = learned_siftLite(im2);
+
+    [matches] = siftMatch(desc1, desc2);
+    newImg = getNewImg(myim1, myim2, matches, locs1, locs2);
+    myvideo(:,:,:,i) = newImg;
+    disp(i);
+end
+% save('../data/myvideo_brief.mat','myvideo');
+% load('../data/myvideo_brief.mat');
+% implay(myvideo);
+writer = VideoWriter('..\data\myvideo_SIFT.avi', ...
+                        'Uncompressed AVI');
+writer.FrameRate = video.FrameRate;
+open(writer);
+for i=1:size(myvideo,4)
+   writeVideo(writer,myvideo(:,:,:,i));
+end
+close(writer);

siftMatch.m

@@ -7,7 +7,7 @@
 %									into desc1 and the second column are indices into desc2
 
 if nargin<3
-    ratio = .6;
+    ratio = .8;
 end
 
 % compute the pairwise Hamming distance