def postprocess(image_info, np_pred_boxes, np_pred_confidences, H, options):
pred_anno = al.Annotation()
#pred_anno.imageName = image_info['path']
#pred_anno.imagePath = os.path.abspath(image_info['path'])
_, rects = add_rectangles(H, [image_info['transformed']],
np_pred_confidences,
np_pred_boxes,
use_stitching=True,
rnn_len=H['rnn_len'],
min_conf=options['min_conf'],
tau=options['tau'],
show_suppressed=False)
rects = [
r for r in rects
if r.x1 < r.x2 and r.y1 < r.y2 and r.score > options['min_conf']
]
h, w = image_info['original'].shape[:2]
if 'rotate90' in H['data'] and H['data']['rotate90']:
# original image height is a width for roatated one
rects = Rotate90.invert(h, rects)
pred_anno.rects = rects
pred_anno = rescale_boxes((H['image_height'], H['image_width']), pred_anno,
h, w)
return pred_anno
def hot_predict_img(image, parameters, to_json=True):
"""Makes predictions when all long running preparation operations are made.
Args:
image_path (string): The path to the source image.
parameters (dict): The parameters produced by :func:`initialize`.
Returns (Annotation):
The annotation for the source image.
"""
H = parameters['hypes']
# The default options for prediction of bounding boxes.
options = H['evaluate']
if 'pred_options' in parameters:
# The new options for prediction of bounding boxes
for key, val in parameters['pred_options'].items():
options[key] = val
# predict
orig_img = (image)[:, :, :3]
img = Rotate90.do(
orig_img
) if 'rotate90' in H['data'] and H['data']['rotate90'] else orig_img
img = imresize(img, (512, 512), interp='cubic')
np_pred_boxes, np_pred_confidences = parameters['sess']. \
run([parameters['pred_boxes'], parameters['pred_confidences']],
feed_dict={parameters['x_in']: img})
image_info = {'original': orig_img, 'transformed': img}
pred_anno = postprocess(image_info, np_pred_boxes, np_pred_confidences, H,
options)
result = [r.writeJSON() for r in pred_anno] if to_json else pred_anno
return result
def preprocess_image(anno, H):
image = imread(anno.imageName)
# skip greyscale images
if len(image.shape) < 3:
return []
# 4 channels images check
if image.shape[2] == 4:
image = image[:, :, :3]
if 'rotate90' in H['data'] and H['data']['rotate90']:
image, anno = Rotate90.do(image, anno)
if image.shape[0] != H["image_height"] or image.shape[1] != H["image_width"]:
anno = rescale_boxes(image.shape, anno, H["image_height"], H["image_width"])
image = imresize(image, (H["image_height"], H["image_width"]), interp='cubic')
result = [(image, anno)]
if 'augmentations' in H['data']:
new_res = []
for i, a in result:
augmented = Augmentations(H['data']).process(i, a)
new_res.append(augmented)
result = new_res
return result
def sliding_predict(image_path, parameters, to_json, H, options):
orig_img = imread(image_path)[:, :, :3]
height, width, _ = orig_img.shape
if options.get('verbose', False):
print(width, height)
sl_win_options = H['sliding_predict']
assert (sl_win_options['window_height'] > sl_win_options['overlap'])
slides = propose_slides(height, sl_win_options['window_height'], sl_win_options['overlap'])
result = []
for top, bottom in slides:
bottom = min(height, top + sl_win_options['window_height'])
if options.get('verbose', False):
print('Slide: ', 0, top, width, bottom)
img = orig_img[top:bottom, 0:width]
img = Rotate90.do(img)[0] if 'rotate90' in H['data'] and H['data']['rotate90'] else img
img = imresize(img, (H['image_height'], H['image_width']), interp='cubic')
np_pred_boxes, np_pred_confidences = parameters['sess']. \
run([parameters['pred_boxes'], parameters['pred_confidences']], feed_dict={parameters['x_in']: img})
image_info = {'path': image_path, 'original_shape': (bottom-top, width), 'transformed': img, 'a': orig_img[top:bottom, 0:width]}
pred_boxes = postprocess_regular(image_info, np_pred_boxes, np_pred_confidences, H, options)
shift_boxes(pred_boxes, top)
result.extend(pred_boxes)
result = combine_boxes(result, sl_win_options['iou_min'], sl_win_options['nms'])
result = [r.writeJSON() for r in result] if to_json else result
return result
def regular_predict(image_path, parameters, to_json, H, options):
orig_img = imread(image_path)[:, :, :3]
img = Rotate90.do(orig_img)[0] if 'rotate90' in H['data'] and H['data']['rotate90'] else orig_img
img = imresize(img, (H['image_height'], H['image_width']), interp='cubic')
np_pred_boxes, np_pred_confidences = parameters['sess']. \
run([parameters['pred_boxes'], parameters['pred_confidences']], feed_dict={parameters['x_in']: img})
image_info = {'path': image_path, 'original_shape': img.shape[:2], 'transformed': img}
pred_anno = postprocess_regular(image_info, np_pred_boxes, np_pred_confidences, H, options)
result = [r.writeJSON() for r in pred_anno] if to_json else pred_anno
return result
def postprocess_regular(image_info, np_pred_boxes, np_pred_confidences, H, options):
pred_anno = al.Annotation()
pred_anno.imageName = image_info['path']
pred_anno.imagePath = os.path.abspath(image_info['path'])
_, rects = add_rectangles(H, [image_info['transformed']], np_pred_confidences, np_pred_boxes, use_stitching=True,
rnn_len=H['rnn_len'], min_conf=options['min_conf'], tau=options['tau'],
show_suppressed=False)
h, w = image_info['original_shape']
if 'rotate90' in H['data'] and H['data']['rotate90']:
# original image height is a width for rotated one
rects = Rotate90.invert(h, rects)
rects = [r for r in rects if r.x1 < r.x2 and r.y1 < r.y2 and r.score > options['min_conf']]
pred_anno.rects = rects
pred_anno = rescale_boxes((H['image_height'], H['image_width']), pred_anno, h, w)
return pred_anno