def random_transform_group_entry(self, image, annotations, transform=None):
""" Randomly transforms image and annotation.
"""
# randomly transform both image and annotations
if transform or self.transform_generator:
if transform is None:
transform = adjust_transform_for_image(
next(self.transform_generator), image,
self.transform_parameters.relative_translation)
# apply transformation to image
image = apply_transform(transform, image,
self.transform_parameters)
# randomly transform the masks and expand so to have a fake channel dimension
for i, mask in enumerate(annotations['masks']):
annotations['masks'][i] = apply_transform(
transform, mask, self.transform_parameters)
annotations['masks'][i] = np.expand_dims(
annotations['masks'][i], axis=2)
# Transform the bounding boxes in the annotations.
annotations['bboxes'] = annotations['bboxes'].copy()
for index in range(annotations['bboxes'].shape[0]):
annotations['bboxes'][index, :] = transform_aabb(
transform, annotations['bboxes'][index, :])
return image, annotations
def random_transform_group_entry(self, image, annotations, masks):
# randomly transform both image and annotations
if self.transform_generator:
transform = adjust_transform_for_image(next(self.transform_generator), image, self.transform_parameters.relative_translation)
image = apply_transform(transform, image, self.transform_parameters)
# randomly transform the masks and expand so to have a fake channel dimension
for m in range(len(masks)):
masks[m] = apply_transform(transform, masks[m], self.transform_parameters)
masks[m] = np.expand_dims(masks[m], axis=2)
# randomly transform the bounding boxes
annotations = annotations.copy()
for index in range(annotations.shape[0]):
annotations[index, :4] = transform_aabb(transform, annotations[index, :4])
return image, annotations, masks
def random_transform_group_entry(self, image, annotations, transform=None):
""" Randomly transforms image and annotation.
"""
# randomly transform both image and annotations
if transform is not None or self.transform_generator:
if transform is None:
transform = adjust_transform_for_image(next(self.transform_generator), image,
self.transform_parameters.relative_translation)
# apply transformation to image
image = apply_transform(transform, image, self.transform_parameters)
# Transform the bounding boxes in the annotations.
annotations['bboxes'] = annotations['bboxes'].copy()
for index in range(annotations['bboxes'].shape[0]):
annotations['bboxes'][index, :] = transform_aabb(transform, annotations['bboxes'][index, :])
return image, annotations
from keras_retinanet.utils.transform import random_transform
from keras_retinanet.utils.image import apply_transform, TransformParameters, adjust_transform_for_image, read_image_bgr
from matplotlib.image import imsave
if __name__ == "__main__":
img = read_image_bgr('./TestFile/foo.jpg')
transform = random_transform(flip_x_chance=1, flip_y_chance=0)
print(transform)
transform_parameters = TransformParameters()
transform = adjust_transform_for_image(
transform, img, transform_parameters.relative_translation)
img_transformed = apply_transform(transform, img, transform_parameters)
imsave('./TestFile/foo_transformed.jpg', img_transformed)
exit(0)
def evaluate_image(model,
img_path,
save_path,
labels_to_names={
0: "no fracture",
1: "fracture"
},
true_coords=None,
IoU_threshold=0.8,
score_threshold=0.6,
transformation_matrix=None,
crop_type=0):
image = read_image_bgr(img_path)
draw = image.copy()
draw = cv2.cvtColor(draw, cv2.COLOR_BGR2RGB)
image = preprocess_image(image)
image, scale = resize_image(image)
if transformation_matrix is not None:
image = apply_transform(transformation_matrix, image, None)
if true_coords is not None:
true_coords = transform_aabb(transformation_matrix, true_coords)
#TODO
if crop_type > 0:
image, true_coords = crop_image(image, true_coords, crop_type)
start = time.time()
boxes, scores, labels = model.predict_on_batch(
np.expand_dims(image, axis=0))
#print("processing time: ", time.time()-start)
boxes /= scale
num_truth_w_threshold = 0
num_pred_w_threshold = 0
pos_sensitivity = 0
num_true_coords = 0
if true_coords is not None:
coords_copy = true_coords.copy(
) # allows removing truth boxes that already have an associated prediction
for box, score, label in zip(boxes[0], scores[0], labels[0]):
if score < score_threshold:
break
#color = label_color(label)
color = (255, 0, 0) # red for bad prediction
for i in range(len(coords_copy)):
coords = coords_copy[i]
if IoU(box, coords) > IoU_threshold:
print("Prediction: " + str(box) + "; Truth: " + str(coords) +
"; Score: " + str(score) + "; IoU: " +
str(IoU(box, coords)))
num_truth_w_threshold += 1
coords_copy = np.delete(coords_copy, i, 0)
print(coords_copy)
color = (0, 255, 0) # green for good prediction
break
b = box.astype(int)
draw_box(draw, b, color=color)
#caption = "{}{:.3f}".format(labels_to_names[label], score)
#draw_caption(draw, b, caption)
num_pred_w_threshold += 1
for coords in true_coords:
draw_box(draw, coords, color=(0, 0, 255)) # blue for true
pos_sensitivity = num_truth_w_threshold / len(true_coords)
num_true_coords = len(true_coords)
plt.figure(figsize=(15, 15))
plt.axis('off')
plt.imshow(draw)
plt.savefig(save_path + ".jpg")
#plt.show()
plt.close()
try:
pos_pred_val = num_truth_w_threshold / num_pred_w_threshold
except ZeroDivisionError:
pos_pred_val = 0
return pos_sensitivity, pos_pred_val, num_truth_w_threshold, num_true_coords, num_pred_w_threshold
