def get_hidden_detections(sess,
H,
hidden_x_in,
hidden_pred_boxes,
hidden_pred_confidences,
page_images: List[np.ndarray],
crop_whitespace: bool = True,
conf_threshold: float = .5) -> List[List[BoxClass]]:
input_shape = [H['image_height'], H['image_width'], H['image_channels']]
page_datas = [{
'page_image':
page_image,
'orig_size':
page_image.shape[:2],
'resized_page_image':
image_util.imresize_multichannel(page_image, input_shape),
} for page_image in page_images]
predictions = [
sess.run([hidden_pred_boxes, hidden_pred_confidences],
feed_dict={hidden_x_in: page_data['resized_page_image']})
for page_data in page_datas
]
for (page_data, prediction) in zip(page_datas, predictions):
(np_pred_boxes, np_pred_confidences) = prediction
new_img, rects = train_utils.add_rectangles(
H,
page_data['resized_page_image'],
np_pred_confidences,
np_pred_boxes,
use_stitching=True,
min_conf=conf_threshold,
show_suppressed=False)
detected_boxes = [
BoxClass(x1=r.x1, y1=r.y1, x2=r.x2,
y2=r.y2).resize_by_page(input_shape,
page_data['orig_size'])
for r in rects if r.score > conf_threshold
]
if crop_whitespace:
detected_boxes = [
box.crop_whitespace_edges(page_data['page_image'])
for box in detected_boxes
]
detected_boxes = list(filter(None, detected_boxes))
page_data['detected_boxes'] = detected_boxes
return [page_data['detected_boxes'] for page_data in page_datas]
def evaluate_dataset_on_weights(hidden_set_dir: str, hidden_set_images_subdir: str, save_dir: str, iteration: int):
detector_args = {
'save_dir': save_dir,
'iteration': iteration
}
detector = get_detector(detector_args=detector_args)
annos = json.load(open(os.path.join(hidden_set_dir, 'figure_boundaries.json')))
for anno in annos:
image_np = imageio.imread(os.path.join(hidden_set_dir, hidden_set_images_subdir, anno['image_path']))
pred_boxes = detector.get_detections([image_np])[0]
true_boxes = [BoxClass(x1=rect['x1'], y1=rect['y1'], x2=rect['x2'], y2=rect['y2']) for rect in anno['rects']]
(pred_indices, true_indices) = figure_utils.pair_boxes(pred_boxes, true_boxes)
print("Pred boxes: ", pred_boxes)
print("True boxes: ", true_boxes)
print("Pred indices: ", pred_indices)
print("True indices: ", true_indices)
a = 0
def rect_to_box_class(rect: Dict) -> BoxClass:
return BoxClass(x1=rect['x1'], y1=rect['y1'], x2=rect['x2'], y2=rect['y2'])