def _get(self):
net_size = self._net_size
# 1. get input file & its annotation
fname, boxes, coded_labels = parse_annotation(
self.ann_fnames[self._index], self.img_dir, self.lable_names)
# 2. read image in fixed size
img_augmenter = ImgAugment(net_size, net_size, self.jitter)
img, boxes_ = img_augmenter.imread(fname, boxes)
# 3. Append ys
list_ys = _create_empty_xy(net_size, len(self.lable_names))
for original_box, label in zip(boxes_, coded_labels):
max_anchor, scale_index, box_index = _find_match_anchor(
original_box, self.anchors)
_coded_box = _encode_box(list_ys[scale_index], original_box,
max_anchor, net_size, net_size)
_assign_box(list_ys[scale_index], box_index, _coded_box, label)
self._index += 1
if self._index == len(self.ann_fnames):
self._index = 0
self._end_epoch = True
return normalize(img), list_ys[2], list_ys[1], list_ys[0]
def get(self, i):
index = i % len(self.ann_fnames)
fname, boxes, coded_labels = parse_annotation(self.ann_fnames[index],
self.img_dir,
self.lable_names)
# net_size = self._get_net_size(idx)
net_size = self.min_net_size
list_ys = _create_empty_xy(net_size, len(self.lable_names))
# 1. get input file & its annotation
# 2. read image in fixed size
img_augmenter = ImgAugment(net_size, net_size, self.jitter)
img, boxes_ = img_augmenter.imread(fname, boxes)
# 4. Append ys
for original_box, label in zip(boxes_, coded_labels):
max_anchor, scale_index, box_index = _find_match_anchor(
original_box, self.anchors)
_coded_box = _encode_box(list_ys[scale_index], original_box,
max_anchor, net_size, net_size)
_assign_box(list_ys[scale_index], box_index, _coded_box, label)
return normalize(img), list_ys[2], list_ys[1], list_ys[0]
def _parseAnnotation(annFile, labelNames):
itemDir = os.path.split(annFile)[0]
imageFile, boxes, labels = parse_annotation(annFile, itemDir,
labelNames)
# preserve only desired objects (boxes with required labels)
boxes, labels = unzip((b, l) for b, l in zip(boxes, labels) if l > -1)
return imageFile, boxes, labels
def create_generator(image_dir, annotation_dir):
from yolo.dataset.annotation import parse_annotation
train_anns = parse_annotation(annotation_dir,
image_dir,
labels_naming=["raccoon"])
generator = BatchGenerator(train_anns,
anchors=[
17, 18, 28, 24, 36, 34, 42, 44, 56, 51, 72,
66, 90, 95, 92, 154, 139, 281
],
min_net_size=288,
max_net_size=288,
shuffle=False)
return generator
def run(self, threshold=0.5, save_dname=None):
n_true_positives = 0
n_truth = 0
n_pred = 0
for ann_fname in tqdm(self._ann_fnames):
img_fname, true_boxes, true_labels = parse_annotation(
ann_fname, self._img_dname, self._cls_labels)
true_labels = np.array(true_labels)
image = cv2.imread(img_fname)[:, :, ::-1]
boxes, labels, probs = self._detector.detect(image, threshold)
n_true_positives += count_true_positives(boxes, true_boxes, labels,
true_labels)
n_truth += len(true_boxes)
n_pred += len(boxes)
if save_dname:
self._save_img(save_dname, img_fname, image, boxes, labels,
probs)
return calc_score(n_true_positives, n_truth, n_pred)
def copyImageAndAnnotations(annotationFiles, imagesDir, dstDir):
recreateDir(dstDir)
for annotationFile in annotationFiles:
imgFile, _, _ = parse_annotation(annotationFile, imagesDir, [])
shutil.copy(imgFile, dstDir)
shutil.copy(annotationFile, dstDir)