def setUp(self):
# sample data generator
batch_size = 64
self.dataset = DetectionDataset(data_type='train')
self.imgs, self.labs_info = self.dataset[:batch_size]
print('Number of Classes : {}'.format(self.dataset.num_classes))
print('Image shape : {} || Label shape : {}'.format(self.imgs.shape, self.labs_info.shape))
self.strides = [4, 8, 16]
self.scales = [10, 25, 40]
self.ratios = [(1, 1),
(1.5, 0.5),
(1.2, 0.8),
(0.8, 1.2),
(1.4, 1.4)]
self.prior = PriorBoxes(self.strides, self.scales, self.ratios)
self.prior_boxes = self.prior.generate((128, 128)) # prior boxes shape : (6720, 4)
# 0번째 이미지를 쌤플 이미지로 사용함.
self.group_labs = self.labs_info.groupby('image_index')
for ind, labs in self.group_labs:
self.labs = labs
break
self.gt_boxes = self.labs[['cx', 'cy', 'w', 'h']].values
self.gt_labels = self.labs['label'].values
self.iou = calculate_iou(self.prior_boxes, self.gt_boxes)
print('Ground Truths Shape : {}'.format(self.gt_boxes.shape))
print('IOU Shape : {}'.format(self.iou.shape))
print(list(self.labs.groupby('image_index')))
def label_generator(train_labels_bucket, prior_boxes, n_classes):
train_labels = []
for index, train_label_df in train_labels_bucket:
gt_boxes = train_label_df[['cx', 'cy', 'w', 'h']].values # shape => [N_prior , 4]
gt_labels = train_label_df['label'].values # shape => [N_prior, N_classes]
# Calculate IOU
iou = calculate_iou(prior_boxes, gt_boxes)
# Generate Detection Labels
matched_labels = matching_labels_coords(prior_boxes, iou, n_classes, gt_boxes, gt_labels)
train_labels.append(matched_labels)
return np.asarray(train_labels)
for (x, y, w, h) in rects:
proposedRects.append((x, y, x + w, y + h))
# to count number of positive and negative region of interests saved
positiveROIs = 0
negativeROIs = 0
# loop over max number of region proposals
for proposedRect in proposedRects[:config.MAX_PROPOSALS_NO]:
# unpack the proposed rectangle bounding box
(proposedStartX, proposedStartY, proposedEndX,
proposedEndY) = proposedRect
# loop over the ground-truth bounding boxes
for boundBox in boundBoxes:
iou = iou_fn.calculate_iou(boundBox, proposedRect)
(boundStartX, boundStartY, boundEndX, boundEndY) = boundBox
roi = None
outputPath = None
if iou > 0.7 and positiveROIs <= config.MAX_POSITIVE_NO:
# extract ROI and derive positive output path
roi = image[proposedStartY:proposedEndY,
proposedStartX:proposedEndX]
filename = "{}.png".format(totalPositive)
outputPath = os.path.sep.join([config.POSITVE_PATH, filename])
# increment positive counters
positiveROIs += 1
totalPositive += 1