| function LM2OpenCV(database, HOMEIMAGES, base_folder, num_neg, patch_size) | |
| % LM2OpenCV Outputs query in OpenCV Haar-detector format. | |
| % LM2OpenCV(database, HOMEIMAGES, base_folder, num_neg, patch_size) | |
| % converts and stores the object images specified by the LabelMe database | |
| % struct to the OpenCV haar-detector format. base_folder specifies the | |
| % target directory for which positive and negative training examples are | |
| % stored for input to the OpenCV haar-detector training utility | |
| % (createsamples.exe and haartraining.exe). | |
| % | |
| % To form positive training examples, object images are copied from the | |
| % LabelMe database into the folder base_folder/positives and the file | |
| % positives.txt is created and stored in base_folder. The positives.txt | |
| % file lists for each image in database the object bounding boxes and is | |
| % provided as input to the OpenCV createsamples routine. | |
| % | |
| % The negative examples are generated by taking num_neg subwindows of | |
| % size patch_size=[width height] from the images in database that do not | |
| % contain an object instance. These windows are sampled at random | |
| % locations in the images. The cropped subwindows are stored in the | |
| % directory base_folder/negatives and the file negatives.txt is created | |
| % and stored in base_folder, which lists the image filenames stored in | |
| % the negatives directory. The negatives.txt file is used by the OpenCV | |
| % haartraining routine. | |
| % | |
| % The variable HOMEIMAGES specifies the physical location of the LabelMe | |
| % database used by the database index database. | |
| % | |
| % To understand how to train an OpenCV haar-detector with the training | |
| % samples and files generated by this script use example command line | |
| % calls are provided below. | |
| % | |
| % Example (Pedestrian Detector): | |
| % | |
| % (In MATLAB) | |
| % | |
| % HOMEIMAGES = 'C:/LabelMe/Images'; | |
| % HOMEANNOTATIONS = 'C:/LabelMe/Annotations'; | |
| % base_folder = 'D:/dtd'; | |
| % num_neg = 10000; | |
| % patch_size = [16 32]; % width=16, height=32 | |
| % | |
| % LMdb = LMdatabase(HOMEANNOTATIONS); | |
| % query = 'pedestrian,person,human,man,woman'; | |
| % database = LMquery(LMdb, 'object.name', query); | |
| % LM2OpenCV(database, HOMEIMAGES, base_folder, num_neg, patch_size); | |
| % counts = LMcountobject(database, query); | |
| % | |
| % (Command Prompt, Let counts=900) | |
| % | |
| % createsamples -info D:\dtd\positive.txt -vec D:\dtd\positives.vec | |
| % -num 900 -w 16 -h 32 | |
| % | |
| % haartraining -data D:\dtd\peddetector\ -vec D:\dtd\positives.vec -bg | |
| % D:\dtd\negatives.txt -npos 900 -nneg 10000 -w 16 -h 32 | |
| % | |
| % In the above example, the call to createsamples generates the file | |
| % positives.vec used by the haartraining routine. The haartraining | |
| % routine then creates the OpenCV Haar pedestrian detector and saves it | |
| % in the director 'D:\dtd\peddetector'. | |
| % | |
| % get number of images in the database | |
| Nimages = length(database); | |
| % sample negatives evenly from each image | |
| nneg_per_image = ceil(num_neg/Nimages); | |
| if(nneg_per_image<1) | |
| nneg_per_image = 1; | |
| end | |
| % create directories positives and negatives | |
| pos_res = mkdir(base_folder, 'positives'); | |
| neg_res = mkdir(base_folder, 'negatives'); | |
| if ~(pos_res & neg_res) | |
| error('LM2OpenCV error: Unable to create positives and/or negatives directories.'); | |
| end | |
| % create files positives.txt and negatives.txt | |
| pbdir = sprintf('%s/positives',base_folder); | |
| nbdir = sprintf('%s/negatives',base_folder); | |
| pfilename = sprintf('%s/positives.txt', base_folder); | |
| nfilename = sprintf('%s/negatives.txt', base_folder); | |
| pos_file = fopen(pfilename, 'w+t'); | |
| neg_file = fopen(nfilename, 'w+t'); | |
| if (pos_file<0 || neg_file<0) | |
| error('LM2OpenCV error: Unable to create files positive.txt and/or negatives.txt.'); | |
| end | |
| % traverse images: | |
| % - copy over positive image files to the positives directory | |
| % - crop negative examples and store in negatives directory | |
| % - create files positives.txt and negatives.txt | |
| neg_ctr = 0; | |
| obj_ctr = 0; | |
| pw = patch_size(1); | |
| ph = patch_size(2); | |
| for i = 1:Nimages | |
| if isfield(database(i).annotation, 'object') | |
| Nobjects = length(database(i).annotation.object); | |
| try | |
| % load image | |
| img = LMimread(database, i, HOMEIMAGES); % Load image | |
| [nrows ncols c] = size(img); | |
| if (c==3) % convert to grayscale? (I'm not sure if this is required by OpenCV...) | |
| img = rgb2gray(img); | |
| end; | |
| % crop and add positive examples from image i to positive image | |
| % directory | |
| bboxes = []; % store bounding boxes for negative patch generation | |
| for j = 1:Nobjects | |
| [X,Y] = getLMpolygon(database(i).annotation.object(j).polygon); | |
| % compute bounding box of object | |
| x = min(X)-2; | |
| y = min(Y)-2; | |
| w = max(X)+2 - x; | |
| h = max(Y)+2 - y; | |
| ctr_x = x+w/2; | |
| ctr_y = y+h/2; | |
| % round width and height to be a multiple of the patch size | |
| % width and height, and re-center bounding box | |
| sfactor = max(w/pw,h/ph); | |
| w = floor(sfactor*pw); | |
| h = floor(sfactor*ph); | |
| x = floor(ctr_x - w/2); | |
| y = floor(ctr_y - h/2); | |
| % make sure that patch fits inside image (otherwise skip | |
| % this positive) | |
| if (w>ncols || h>nrows) | |
| continue; | |
| end | |
| % check boundaries | |
| if(x<1); x = 1; end; | |
| if(y<1); y = 1; end; | |
| if((x+w)>ncols); x = ncols-w; end; | |
| if((y+h)>nrows); y = nrows-h; end; | |
| % crop and save image in positive image directory | |
| img_filename = sprintf('%s/pos%06d.jpg', pbdir, obj_ctr); | |
| cimg = img(y:(y+h-1),x:(x+w-1)); | |
| imwrite(cimg, img_filename, 'JPG', 'Quality', 100); | |
| % add entry into positives file: | |
| % OpenCV format is upper-left corner (x,y), width (w) and | |
| % height (h) | |
| fprintf(pos_file, 'positives\\pos%06d.jpg\t1\t0 0 %d %d\n', obj_ctr, w, h); | |
| % use bboxes below... | |
| bboxes = [bboxes; x y w h]; | |
| % next object | |
| obj_ctr = obj_ctr + 1; | |
| end | |
| % generate negative patches | |
| x = []; | |
| y = []; | |
| for j = 1:nneg_per_image | |
| if neg_ctr > num_neg | |
| break; | |
| end | |
| cx = floor(rand(1)*(ncols-pw-1)); | |
| cy = floor(rand(1)*(nrows-ph-1)); | |
| % check boundaries | |
| if(cx<1); cx = 1; end; | |
| if(cy<1); cy = 1; end; | |
| if((cx+pw)>ncols); cx = ncols-pw; end; | |
| if((cy+ph)>nrows); cy = nrows-ph; end; | |
| while (sum(cx==x) || sum(cy==y) || ... | |
| isOverlap(cx,cy,pw,ph,bboxes)) | |
| cx = floor(rand(1)*(ncols-pw-1)); | |
| cy = floor(rand(1)*(nrows-ph-1)); | |
| % check boundaries | |
| if(cx<1); cx = 1; end; | |
| if(cy<1); cy = 1; end; | |
| if((cx+pw)>ncols); cx = ncols-pw; end; | |
| if((cy+ph)>nrows); cy = nrows-ph; end; | |
| end | |
| x = [x cx]; | |
| y = [y cy]; | |
| % crop and save negative patch | |
| cimg = img(cy:(cy+ph-1),cx:(cx+pw-1)); | |
| img_filename = sprintf('%s/neg%06d.jpg', nbdir, neg_ctr); | |
| imwrite(cimg, img_filename, 'JPG', 'Quality', 100); | |
| % update negatives file | |
| fprintf(neg_file, 'negatives\\neg%06d.jpg\n', neg_ctr); | |
| neg_ctr = neg_ctr + 1; | |
| end | |
| catch | |
| i | |
| 'dimensions (x, y, w, h)' | |
| x | |
| y | |
| w | |
| h | |
| end | |
| end | |
| end | |
| % done. | |
| fclose(pos_file); | |
| fclose(neg_file); | |
| %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% | |
| % isOverlap | |
| % | |
| function ans = isOverlap(x, y, w, h, bboxes) | |
| ans = 0; | |
| n = size(bboxes,1); | |
| for idx = 1:n | |
| cx = bboxes(idx,1); | |
| cy = bboxes(idx,2); | |
| cw = bboxes(idx,3); | |
| ch = bboxes(idx,4); | |
| if ~(x>(cx+cw) || y>(cy+ch) || (x+w)<cx || (y+h)<cy) | |
| ans = 1; | |
| return; | |
| end | |
| end | |