def visualize_output_rois(testing=False):
p = PARAMETERS.get_parameters_for_dataset()
# no need to change these parameters
boUseNonMaximaSurpression = True
visualizationDir = os.path.join(p.resultsDir, "visualizations")
cntkParsedOutputDir = os.path.join(p.cntkFilesDir, image_set + "_parsed")
makeDirectory(p.resultsDir)
makeDirectory(visualizationDir)
# loop over all images and visualize
imdb = p.imdbs[image_set]
for imgIndex in range(0, imdb.num_images):
imgPath = imdb.image_path_at(imgIndex)
imgWidth, imgHeight = imWidthHeight(imgPath)
# evaluate classifier for all rois
labels, scores = nnPredict(imgIndex, cntkParsedOutputDir,
p.cntk_nrRois, len(p.classes), None)
# remove the zero-padded rois
scores = scores[:len(imdb.roidb[imgIndex]['boxes'])]
labels = labels[:len(imdb.roidb[imgIndex]['boxes'])]
# perform non-maxima surpression. note that the detected classes in the image is not affected by this.
nmsKeepIndices = []
if boUseNonMaximaSurpression:
nmsKeepIndices = applyNonMaximaSuppression(
p.nmsThreshold, labels, scores, imdb.roidb[imgIndex]['boxes'])
print(
"Non-maxima surpression kept {:4} of {:4} rois (nmsThreshold={})"
.format(len(nmsKeepIndices), len(labels), p.nmsThreshold))
# visualize results
imgDebug = visualizeResults(imgPath,
labels,
scores,
imdb.roidb[imgIndex]['boxes'],
p.cntk_padWidth,
p.cntk_padHeight,
p.classes,
nmsKeepIndices,
boDrawNegativeRois=True)
if not testing:
imshow(imgDebug, waitDuration=0, maxDim=800)
# imwrite(imgDebug, visualizationDir + "/" + str(imgIndex) + os.path.basename(imgPath))
print("DONE.")
return True
def evaluate_rois():
p = PARAMETERS.get_parameters_for_dataset()
overlaps = []
roiCounts = []
for subdir in subdirs:
imgFilenames = getFilesInDirectory(os.path.join(p.imgDir, subdir),
".jpg")
# loop over all iamges
for imgIndex, imgFilename in enumerate(imgFilenames):
if imgIndex % 20 == 0:
print("Processing subdir '{}', image {} of {}".format(
subdir, imgIndex, len(imgFilenames)))
# load ground truth
imgPath = os.path.join(p.imgDir, subdir, imgFilename)
imgWidth, imgHeight = imWidthHeight(imgPath)
gtBoxes, gtLabels = readGtAnnotation(imgPath)
gtBoxes = [Bbox(*rect) for rect in gtBoxes]
# load rois and compute scale
rois = readRois(p.roiDir, subdir, imgFilename)
rois = [Bbox(*roi) for roi in rois]
roiCounts.append(len(rois))
# for each ground truth, compute if it is covered by an roi
maxOverlap = -1
for gtIndex, (gtLabel, gtBox) in enumerate(zip(gtLabels, gtBoxes)):
assert (gtBox.max() <= max(imgWidth, imgHeight)
and gtBox.max() >= 0)
gtLabel = gtLabel.decode('utf-8')
if gtLabel in p.classes[1:]:
for roi in rois:
assert (roi.max() <= max(imgWidth, imgHeight)
and roi.max() >= 0)
overlap = bboxComputeOverlapVoc(gtBox, roi)
maxOverlap = max(maxOverlap, overlap)
overlaps.append(maxOverlap)
print("Average number of rois per image " +
str(1.0 * sum(roiCounts) / len(overlaps)))
# compute recall at different overlaps
overlaps = np.array(overlaps, np.float32)
for overlapThreshold in np.linspace(0, 1, 11):
recall = 1.0 * sum(overlaps >= overlapThreshold) / len(overlaps)
print("At threshold {:.2f}: recall = {:2.2f}".format(
overlapThreshold, recall))
return True
def evaluate_rois():
p = PARAMETERS.get_parameters_for_dataset()
overlaps = []
roiCounts = []
for subdir in subdirs:
imgFilenames = getFilesInDirectory(os.path.join(p.imgDir, subdir), ".jpg")
# loop over all iamges
for imgIndex,imgFilename in enumerate(imgFilenames):
if imgIndex % 20 == 0:
print ("Processing subdir '{}', image {} of {}".format(subdir, imgIndex, len(imgFilenames)))
# load ground truth
imgPath = os.path.join(p.imgDir, subdir, imgFilename)
imgWidth, imgHeight = imWidthHeight(imgPath)
gtBoxes, gtLabels = readGtAnnotation(imgPath)
gtBoxes = [Bbox(*rect) for rect in gtBoxes]
# load rois and compute scale
rois = readRois(p.roiDir, subdir, imgFilename)
rois = [Bbox(*roi) for roi in rois]
roiCounts.append(len(rois))
# for each ground truth, compute if it is covered by an roi
maxOverlap = -1
for gtIndex, (gtLabel, gtBox) in enumerate(zip(gtLabels,gtBoxes)):
assert (gtBox.max() <= max(imgWidth, imgHeight) and gtBox.max() >= 0)
gtLabel = gtLabel.decode('utf-8')
if gtLabel in p.classes[1:]:
for roi in rois:
assert (roi.max() <= max(imgWidth, imgHeight) and roi.max() >= 0)
overlap = bboxComputeOverlapVoc(gtBox, roi)
maxOverlap = max(maxOverlap, overlap)
overlaps.append(maxOverlap)
print ("Average number of rois per image " + str(1.0 * sum(roiCounts) / len(overlaps)))
# compute recall at different overlaps
overlaps = np.array(overlaps, np.float32)
for overlapThreshold in np.linspace(0,1,11):
recall = 1.0 * sum(overlaps >= overlapThreshold) / len(overlaps)
print ("At threshold {:.2f}: recall = {:2.2f}".format(overlapThreshold, recall))
return True
def visualize_output_rois(testing=False):
p = PARAMETERS.get_parameters_for_dataset()
# no need to change these parameters
boUseNonMaximaSurpression = True
visualizationDir = os.path.join(p.resultsDir, "visualizations")
cntkParsedOutputDir = os.path.join(p.cntkFilesDir, image_set + "_parsed")
makeDirectory(p.resultsDir)
makeDirectory(visualizationDir)
# loop over all images and visualize
imdb = p.imdbs[image_set]
for imgIndex in range(0, imdb.num_images):
imgPath = imdb.image_path_at(imgIndex)
imgWidth, imgHeight = imWidthHeight(imgPath)
# evaluate classifier for all rois
labels, scores = nnPredict(imgIndex, cntkParsedOutputDir, p.cntk_nrRois, len(p.classes), None)
# remove the zero-padded rois
scores = scores[:len(imdb.roidb[imgIndex]['boxes'])]
labels = labels[:len(imdb.roidb[imgIndex]['boxes'])]
# perform non-maxima surpression. note that the detected classes in the image is not affected by this.
nmsKeepIndices = []
if boUseNonMaximaSurpression:
nmsKeepIndices = applyNonMaximaSuppression(p.nmsThreshold, labels, scores, imdb.roidb[imgIndex]['boxes'])
print ("Non-maxima surpression kept {:4} of {:4} rois (nmsThreshold={})".format(len(nmsKeepIndices), len(labels), p.nmsThreshold))
# visualize results
imgDebug = visualizeResults(imgPath, labels, scores, imdb.roidb[imgIndex]['boxes'], p.cntk_padWidth, p.cntk_padHeight,
p.classes, nmsKeepIndices, boDrawNegativeRois=True)
if not testing:
imshow(imgDebug, waitDuration=0, maxDim = 800)
# imwrite(imgDebug, visualizationDir + "/" + str(imgIndex) + os.path.basename(imgPath))
print ("DONE.")
return True
def generate_input_rois(testing=False):
p = PARAMETERS.get_parameters_for_dataset()
if not p.datasetName.startswith("pascalVoc"):
# init
makeDirectory(p.roiDir)
roi_minDim = p.roi_minDimRel * p.roi_maxImgDim
roi_maxDim = p.roi_maxDimRel * p.roi_maxImgDim
roi_minNrPixels = p.roi_minNrPixelsRel * p.roi_maxImgDim*p.roi_maxImgDim
roi_maxNrPixels = p.roi_maxNrPixelsRel * p.roi_maxImgDim*p.roi_maxImgDim
for subdir in subDirs:
makeDirectory(os.path.join(p.roiDir, subdir))
imgFilenames = getFilesInDirectory(os.path.join(p.imgDir, subdir), ".jpg")
# loop over all images
for imgIndex,imgFilename in enumerate(imgFilenames):
roiPath = "{}/{}/{}.roi.txt".format(p.roiDir, subdir, imgFilename[:-4])
# load image
print (imgIndex, len(imgFilenames), subdir, imgFilename)
tstart = datetime.datetime.now()
imgPath = os.path.join(p.imgDir, subdir, imgFilename)
imgOrig = imread(imgPath)
if imWidth(imgPath) > imHeight(imgPath):
print (imWidth(imgPath) , imHeight(imgPath))
# get rois
if boAddSelectiveSearchROIs:
print ("Calling selective search..")
rects, img, scale = getSelectiveSearchRois(imgOrig, p.ss_scale, p.ss_sigma, p.ss_minSize, p.roi_maxImgDim) #interpolation=cv2.INTER_AREA
print (" Number of rois detected using selective search: " + str(len(rects)))
else:
rects = []
img, scale = imresizeMaxDim(imgOrig, p.roi_maxImgDim, boUpscale=True, interpolation=cv2.INTER_AREA)
imgWidth, imgHeight = imArrayWidthHeight(img)
# add grid rois
if boAddRoisOnGrid:
rectsGrid = getGridRois(imgWidth, imgHeight, p.grid_nrScales, p.grid_aspectRatios)
print (" Number of rois on grid added: " + str(len(rectsGrid)))
rects += rectsGrid
# run filter
print (" Number of rectangles before filtering = " + str(len(rects)))
rois = filterRois(rects, imgWidth, imgHeight, roi_minNrPixels, roi_maxNrPixels, roi_minDim, roi_maxDim, p.roi_maxAspectRatio)
if len(rois) == 0: #make sure at least one roi returned per image
rois = [[5, 5, imgWidth-5, imgHeight-5]]
print (" Number of rectangles after filtering = " + str(len(rois)))
# scale up to original size and save to disk
# note: each rectangle is in original image format with [x,y,x2,y2]
rois = np.int32(np.array(rois) / scale)
assert (np.min(rois) >= 0)
assert (np.max(rois[:, [0,2]]) < imArrayWidth(imgOrig))
assert (np.max(rois[:, [1,3]]) < imArrayHeight(imgOrig))
np.savetxt(roiPath, rois, fmt='%d')
print (" Time [ms]: " + str((datetime.datetime.now() - tstart).total_seconds() * 1000))
# clear imdb cache and other files
if os.path.exists(p.cntkFilesDir):
assert(p.cntkFilesDir.endswith("cntkFiles"))
if not testing:
userInput = input('--> INPUT: Press "y" to delete directory ' + p.cntkFilesDir + ": ")
if userInput.lower() not in ['y', 'yes']:
print ("User input is %s: exiting now." % userInput)
exit(-1)
shutil.rmtree(p.cntkFilesDir)
time.sleep(0.1) # avoid access problems
# create cntk representation for each image
for image_set in image_sets:
imdb = p.imdbs[image_set]
print ("Number of images in set {} = {}".format(image_set, imdb.num_images))
makeDirectory(p.cntkFilesDir)
# open files for writing
cntkImgsPath, cntkRoiCoordsPath, cntkRoiLabelsPath, nrRoisPath = getCntkInputPaths(p.cntkFilesDir, image_set)
with open(nrRoisPath, 'w') as nrRoisFile, \
open(cntkImgsPath, 'w') as cntkImgsFile, \
open(cntkRoiCoordsPath, 'w') as cntkRoiCoordsFile, \
open(cntkRoiLabelsPath, 'w') as cntkRoiLabelsFile:
# for each image, transform rois etc to cntk format
for imgIndex in range(0, imdb.num_images):
if imgIndex % 50 == 0:
print ("Processing image set '{}', image {} of {}".format(image_set, imgIndex, imdb.num_images))
currBoxes = imdb.roidb[imgIndex]['boxes']
currGtOverlaps = imdb.roidb[imgIndex]['gt_overlaps']
imgPath = imdb.image_path_at(imgIndex)
imgWidth, imgHeight = imWidthHeight(imgPath)
# all rois need to be scaled + padded to cntk input image size
targetw, targeth, w_offset, h_offset, scale = roiTransformPadScaleParams(imgWidth, imgHeight,
p.cntk_padWidth, p.cntk_padHeight)
boxesStr = ""
labelsStr = ""
nrBoxes = len(currBoxes)
for boxIndex, box in enumerate(currBoxes):
rect = roiTransformPadScale(box, w_offset, h_offset, scale)
boxesStr += getCntkRoiCoordsLine(rect, p.cntk_padWidth, p.cntk_padHeight)
labelsStr += getCntkRoiLabelsLine(currGtOverlaps[boxIndex, :].toarray()[0],
p.train_posOverlapThres,
p.nrClasses)
# if less than e.g. 2000 rois per image, then fill in the rest using 'zero-padding'.
boxesStr, labelsStr = cntkPadInputs(nrBoxes, p.cntk_nrRois, p.nrClasses, boxesStr, labelsStr)
# update cntk data
nrRoisFile.write("{}\n".format(nrBoxes))
cntkImgsFile.write("{}\t{}\t0\n".format(imgIndex, imgPath))
cntkRoiCoordsFile.write("{} |rois{}\n".format(imgIndex, boxesStr))
cntkRoiLabelsFile.write("{} |roiLabels{}\n".format(imgIndex, labelsStr))
print ("DONE.")
return True
def generate_input_rois(testing=False):
p = PARAMETERS.get_parameters_for_dataset()
if not p.datasetName.startswith("pascalVoc"):
# init
makeDirectory(p.roiDir)
roi_minDim = p.roi_minDimRel * p.roi_maxImgDim
roi_maxDim = p.roi_maxDimRel * p.roi_maxImgDim
roi_minNrPixels = p.roi_minNrPixelsRel * p.roi_maxImgDim*p.roi_maxImgDim
roi_maxNrPixels = p.roi_maxNrPixelsRel * p.roi_maxImgDim*p.roi_maxImgDim
for subdir in subDirs:
makeDirectory(os.path.join(p.roiDir, subdir))
imgFilenames = getFilesInDirectory(os.path.join(p.imgDir, subdir), ".jpg")
# loop over all images
for imgIndex,imgFilename in enumerate(imgFilenames):
roiPath = "{}/{}/{}.roi.txt".format(p.roiDir, subdir, imgFilename[:-4])
# load image
print (imgIndex, len(imgFilenames), subdir, imgFilename)
tstart = datetime.datetime.now()
imgPath = os.path.join(p.imgDir, subdir, imgFilename)
imgOrig = imread(imgPath)
if imWidth(imgPath) > imHeight(imgPath):
print (imWidth(imgPath) , imHeight(imgPath))
# get rois
if boAddSelectiveSearchROIs:
print ("Calling selective search..")
rects, img, scale = getSelectiveSearchRois(imgOrig, p.ss_scale, p.ss_sigma, p.ss_minSize, p.roi_maxImgDim) #interpolation=cv2.INTER_AREA
print (" Number of rois detected using selective search: " + str(len(rects)))
else:
rects = []
img, scale = imresizeMaxDim(imgOrig, p.roi_maxImgDim, boUpscale=True, interpolation=cv2.INTER_AREA)
imgWidth, imgHeight = imArrayWidthHeight(img)
# add grid rois
if boAddRoisOnGrid:
rectsGrid = getGridRois(imgWidth, imgHeight, p.grid_nrScales, p.grid_aspectRatios)
print (" Number of rois on grid added: " + str(len(rectsGrid)))
rects += rectsGrid
# run filter
print (" Number of rectangles before filtering = " + str(len(rects)))
rois = filterRois(rects, imgWidth, imgHeight, roi_minNrPixels, roi_maxNrPixels, roi_minDim, roi_maxDim, p.roi_maxAspectRatio)
if len(rois) == 0: #make sure at least one roi returned per image
rois = [[5, 5, imgWidth-5, imgHeight-5]]
print (" Number of rectangles after filtering = " + str(len(rois)))
# scale up to original size and save to disk
# note: each rectangle is in original image format with [x,y,x2,y2]
rois = np.int32(np.array(rois) / scale)
assert (np.min(rois) >= 0)
assert (np.max(rois[:, [0,2]]) < imArrayWidth(imgOrig))
assert (np.max(rois[:, [1,3]]) < imArrayHeight(imgOrig))
np.savetxt(roiPath, rois, fmt='%d')
print (" Time [ms]: " + str((datetime.datetime.now() - tstart).total_seconds() * 1000))
# clear imdb cache and other files
if os.path.exists(p.cntkFilesDir):
assert(p.cntkFilesDir.endswith("cntkFiles"))
if not testing:
userInput = input('--> INPUT: Press "y" to delete directory ' + p.cntkFilesDir + ": ")
if userInput.lower() not in ['y', 'yes']:
print ("User input is %s: exiting now." % userInput)
exit(-1)
shutil.rmtree(p.cntkFilesDir)
time.sleep(0.1) # avoid access problems
# create cntk representation for each image
for image_set in image_sets:
imdb = p.imdbs[image_set]
print ("Number of images in set {} = {}".format(image_set, imdb.num_images))
makeDirectory(p.cntkFilesDir)
# open files for writing
cntkImgsPath, cntkRoiCoordsPath, cntkRoiLabelsPath, nrRoisPath = getCntkInputPaths(p.cntkFilesDir, image_set)
with open(nrRoisPath, 'w') as nrRoisFile, \
open(cntkImgsPath, 'w') as cntkImgsFile, \
open(cntkRoiCoordsPath, 'w') as cntkRoiCoordsFile, \
open(cntkRoiLabelsPath, 'w') as cntkRoiLabelsFile:
# for each image, transform rois etc to cntk format
for imgIndex in range(0, imdb.num_images):
if imgIndex % 50 == 0:
print ("Processing image set '{}', image {} of {}".format(image_set, imgIndex, imdb.num_images))
currBoxes = imdb.roidb[imgIndex]['boxes']
currGtOverlaps = imdb.roidb[imgIndex]['gt_overlaps']
imgPath = imdb.image_path_at(imgIndex)
imgWidth, imgHeight = imWidthHeight(imgPath)
# all rois need to be scaled + padded to cntk input image size
targetw, targeth, w_offset, h_offset, scale = roiTransformPadScaleParams(imgWidth, imgHeight,
p.cntk_padWidth, p.cntk_padHeight)
boxesStr = ""
labelsStr = ""
nrBoxes = len(currBoxes)
for boxIndex, box in enumerate(currBoxes):
rect = roiTransformPadScale(box, w_offset, h_offset, scale)
boxesStr += getCntkRoiCoordsLine(rect, p.cntk_padWidth, p.cntk_padHeight)
labelsStr += getCntkRoiLabelsLine(currGtOverlaps[boxIndex, :].toarray()[0],
p.train_posOverlapThres,
p.nrClasses)
# if less than e.g. 2000 rois per image, then fill in the rest using 'zero-padding'.
boxesStr, labelsStr = cntkPadInputs(nrBoxes, p.cntk_nrRois, p.nrClasses, boxesStr, labelsStr)
# update cntk data
nrRoisFile.write("{}\n".format(nrBoxes))
cntkImgsFile.write("{}\t{}\t0\n".format(imgIndex, imgPath))
cntkRoiCoordsFile.write("{} |rois{}\n".format(imgIndex, boxesStr))
cntkRoiLabelsFile.write("{} |roiLabels{}\n".format(imgIndex, labelsStr))
print ("DONE.")
return True