added getDagNNBatch

training/training_code/getDagNNBatch.m

@@ -0,0 +1,45 @@
+function inputs = getDagNNBatch(topts,imagesA,theta,batchIdx)
+
+% specify batch image size (CNN input size)
+imsize = [227 227];
+
+% get images and transformation parameters from the specified batchIdx
+theta_batch = single(theta(:,:,:,batchIdx));
+%im_raw = subMeanAndReshape(vl_imreadjpeg(imagesA(batchIdx))); % images in imagesA should be of the same resolution
+im_raw = subMeanAndReshape(vl_imreadjpeg(imagesA(batchIdx),'resize',[640 480])); % images in imagesA should be of the same resolution
+[H,W]=size(im_raw(:,:,1,1));
+
+% generate image A by cropping the raw image
+innerCropfactor = 9/16;
+imA_batch = im_raw(round(H*(1-innerCropfactor)/2+1):end-round(H*(1-innerCropfactor)/2),round(W*(1-innerCropfactor)/2+1):end-round(W*(1-innerCropfactor)/2),:,:);
+% resize to the batch image size
+imA_batch = imresize(imA_batch,imsize);
+
+% add extra padding for enlarging the sampling region for image B
+paddingFactor = 1/2;
+im_raw = imresize(im_raw,[454 454]);  % delete line
+im_raw=padarray(im_raw, size(im_raw(:,:,1,1))*paddingFactor, 'symmetric');    
+factor = paddingFactor*innerCropfactor;        
+
+% generate image B by transforming image A
+if strcmp(topts.geometricModel,'affine')==1        
+    % use affine transformation
+    tnf = dagnn.AffineGridGenerator('Ho',imsize(1),'Wo',imsize(2)); 
+elseif strcmp(topts.geometricModel,'TPS')==1
+    % use TPS transformation
+    tnf = dagnnExtra.TpsGridGenerator('Ho',imsize(1),'Wo',imsize(2));        
+end
+samplingGrid = tnf.forward({theta_batch});
+bs = dagnn.BilinearSampler;
+imB_batch = bs.forward({im_raw,samplingGrid{1}*factor});
+imB_batch = imB_batch{1};
+
+% copy to GPU memory if needed
+if ~isempty(topts.gpus)
+    imA_batch = gpuArray(imA_batch) ;
+    imB_batch = gpuArray(imB_batch) ;
+    theta_batch =  gpuArray(theta_batch) ;
+end
+
+% return batch data
+inputs = {'thetaGt', theta_batch, 'AN1input', imA_batch, 'AN2input', imB_batch};