def classify(self, img, pixel_rect):
img_h, img_w = img.shape[:2]
img_dims = (img_w, img_h)
w, h = self.window_dims
window_shape = (h, w)
window_aspect = w / float(h)
# print window_aspect
# Enlarge to window_dims:
enlarged_rect = pixel_rect.enlarge_to_aspect(window_aspect)
# Ensure rectangle is located within its image:
object_rect = enlarged_rect.translated([0,0], img_dims)
# if not pixel_rect.lies_within_frame(img_dims):
sample = utils.crop_rectangle(img, object_rect)
sample = utils.resize_sample(sample, shape=window_shape)
sample = self.prepare_sample(sample)
flat_sample = sample.reshape(sample.shape[0] * sample.shape[1] * sample.shape[2])
feed = {self.x: [flat_sample], self.keep_prob: 1.0}
label_probs = self.sess.run(tf.nn.softmax(self.logits), feed_dict=feed)
pos_prob, neg_prob = label_probs[0]
is_object = pos_prob > neg_prob
return is_object, object_rect, pos_prob
def classify(self, img, pixel_rect):
img_h, img_w = img.shape[:2]
img_dims = (img_w, img_h)
w, h = self.window_dims
window_shape = (h, w)
window_aspect = w / float(h)
# print window_aspect
# Enlarge to window_dims:
enlarged_rect = pixel_rect.enlarge_to_aspect(window_aspect)
# Ensure rectangle is located within its image:
object_rect = enlarged_rect.translated([0, 0], img_dims)
# if not pixel_rect.lies_within_frame(img_dims):
sample = utils.crop_rectangle(img, object_rect)
sample = utils.resize_sample(sample, shape=window_shape)
sample = self.prepare_sample(sample)
flat_sample = sample.reshape(sample.shape[0] * sample.shape[1] *
sample.shape[2])
feed = {self.x: [flat_sample], self.keep_prob: 1.0}
label_probs = self.sess.run(tf.nn.softmax(self.logits), feed_dict=feed)
pos_prob, neg_prob = label_probs[0]
is_object = pos_prob > neg_prob
return is_object, object_rect, pos_prob
def sliding_window_generator(img, window_dims=(32, 32), scale_factor=1.1, strides=(8, 8), only_rects=False):
""" Creates an image pyramid, and generates views into each of its levels.
Returns the pixel rectangle corresponding to each image
"""
h, w = img.shape[:2]
orig_dims = (w, h)
sx, sy = strides
ww, wh = window_dims
for img in image_pyramid(img, scale_factor, window_dims):
h, w = img.shape[:2]
img_dims = (w, h)
mx = w - ww
my = h - wh
for x in xrange(0, mx, sx):
for y in xrange(0, my, sy):
# Create the window in the current pyramid level:
local_rect = gm.PixelRectangle.from_opencv_bbox([x, y, ww, wh])
# Crop the window:
window = None
if not only_rects:
window = utils.crop_rectangle(img, local_rect)
# Find the window location in the original image:
orig_rect = local_rect.scale_image(img_dims, orig_dims)
if (not window is None) and window.shape[0] != wh:
print 'ERROR:', (w, h), (x, y), (mx, my), local_rect
yield (window, orig_rect)
def update_display(self):
h, w = self.current_img.shape[:2]
max_dim = float(self.display_width)
if w > max_dim + 50 or h > max_dim + 50:
sx = max_dim / w
sy = max_dim / h
self.current_scale = min(sx, sy)
self.current_img = cv2.resize(self.current_img,
dsize=None,
fx=self.current_scale,
fy=self.current_scale)
clone_img = self.current_img.copy()
# Draw bounding boxes for current image:
for rect in self.get_image_rectangles(self.current_path):
oh, ow = self.original_shape[:2]
rect = rect.scale_image((ow, oh), (w, h))
if self.flipped:
rect = rect.rotated_180((w, h))
tl = tuple(rect.tl)
br = tuple(rect.br)
cv2.rectangle(clone_img, tl, br, (255, 255, 255), 3)
cv2.rectangle(clone_img, tl, br, (255, 0, 127), 2)
# Draw bounding box being edited:
if self.rect_tl and self.rect_br:
cv2.rectangle(clone_img, self.rect_tl, self.rect_br,
(255, 255, 255), 3)
cv2.rectangle(clone_img, self.rect_tl, self.rect_br, (0, 255, 0),
2)
rect = gm.PixelRectangle.fromCorners(self.rect_tl, self.rect_br)
rect = self.convert_to_rect_in_original(rect)
cropped = utils.crop_rectangle(self.original_img, rect)
cw = self.current_img.shape[1]
ch = self.current_img.shape[0]
if rect.w > 5 and rect.h > 5:
if rect.w > rect.h:
preview_width = cw / 5.0
preview_height = preview_width / rect.aspect
else:
preview_height = ch / 4.0
preview_width = preview_height * rect.aspect
new_shape = (max(2, int(round(preview_height))),
max(2, int(round(preview_width))))
preview = utils.resize_sample(cropped,
new_shape,
use_interp=False)
clone_img[0:new_shape[0], 0:new_shape[1], :] = preview
cv2.imshow(self.winName, clone_img)
def update_display(self):
h, w = self.current_img.shape[:2]
max_dim = float(self.display_width)
if w > max_dim + 50 or h > max_dim + 50:
sx = max_dim / w
sy = max_dim / h
self.current_scale = min(sx, sy)
self.current_img = cv2.resize(self.current_img, dsize=None, fx=self.current_scale, fy=self.current_scale)
clone_img = self.current_img.copy()
# Draw bounding boxes for current image:
for rect in self.get_image_rectangles(self.current_path):
oh, ow = self.original_shape[:2]
rect = rect.scale_image((ow, oh), (w,h))
if self.flipped:
rect = rect.rotated_180((w,h))
tl = tuple(rect.tl)
br = tuple(rect.br)
cv2.rectangle(clone_img, tl, br, (255, 255, 255), 3)
cv2.rectangle(clone_img, tl, br, (255, 0, 127), 2)
# Draw bounding box being edited:
if self.rect_tl and self.rect_br:
cv2.rectangle(clone_img, self.rect_tl, self.rect_br, (255, 255, 255), 3)
cv2.rectangle(clone_img, self.rect_tl, self.rect_br, (0, 255, 0), 2)
rect = gm.PixelRectangle.fromCorners(self.rect_tl, self.rect_br)
rect = self.convert_to_rect_in_original(rect)
cropped = utils.crop_rectangle(self.original_img, rect)
cw = self.current_img.shape[1]
ch = self.current_img.shape[0]
if rect.w > 5 and rect.h > 5:
if rect.w > rect.h:
preview_width = cw / 5.0
preview_height = preview_width / rect.aspect
else:
preview_height = ch / 4.0
preview_width = preview_height * rect.aspect
new_shape = (max(2, int(round(preview_height))), max(2, int(round(preview_width))))
preview = utils.resize_sample(cropped, new_shape, use_interp=False)
clone_img[0:new_shape[0],0:new_shape[1],:] = preview
cv2.imshow(self.winName, clone_img)
def sliding_window_generator(img,
window_dims=(32, 32),
scale_factor=1.1,
strides=(8, 8),
only_rects=False):
""" Creates an image pyramid, and generates views into each of its levels.
Returns the pixel rectangle corresponding to each image
"""
h, w = img.shape[:2]
orig_dims = (w, h)
sx, sy = strides
ww, wh = window_dims
for img in image_pyramid(img, scale_factor, window_dims):
h, w = img.shape[:2]
img_dims = (w, h)
mx = w - ww
my = h - wh
for x in xrange(0, mx, sx):
for y in xrange(0, my, sy):
# Create the window in the current pyramid level:
local_rect = gm.PixelRectangle.from_opencv_bbox([x, y, ww, wh])
# Crop the window:
window = None
if not only_rects:
window = utils.crop_rectangle(img, local_rect)
# Find the window location in the original image:
orig_rect = local_rect.scale_image(img_dims, orig_dims)
if (not window is None) and window.shape[0] != wh:
print 'ERROR:', (w, h), (x, y), (mx, my), local_rect
yield (window, orig_rect)
