def __init__(self,
weights_1st_stage,
scale_space_sizes_idxs,
num_win_psz=130,
edge=EDGE,
base_log=BASE_LOG,
min_edge_log=MIN_EDGE_LOG,
edge_log_range=EDGE_LOG_RANGE):
self.filter_tig = FilterTIG()
self.weights_1st_stage = weights_1st_stage
self.filter_tig.update(self.weights_1st_stage)
self.filter_tig.reconstruct(self.weights_1st_stage)
self.scale_space_sizes_idxs = scale_space_sizes_idxs
self.base_log = base_log
self.min_edge_log = min_edge_log
self.edge_log_range = edge_log_range
self.edge = edge
self.num_win_psz = num_win_psz
def __init__(self, weights_1st_stage, scale_space_sizes_idxs, num_win_psz = 130, edge = EDGE, base_log = BASE_LOG, min_edge_log = MIN_EDGE_LOG, edge_log_range = EDGE_LOG_RANGE):
self.filter_tig = FilterTIG()
self.weights_1st_stage = weights_1st_stage
self.filter_tig.update(self.weights_1st_stage)
self.filter_tig.reconstruct(self.weights_1st_stage)
self.scale_space_sizes_idxs = scale_space_sizes_idxs
self.base_log = base_log
self.min_edge_log = min_edge_log
self.edge_log_range = edge_log_range
self.edge = edge
self.num_win_psz = num_win_psz
class FirstStagePrediction(object):
def __init__(self,
weights_1st_stage,
scale_space_sizes_idxs,
num_win_psz=130,
edge=EDGE,
base_log=BASE_LOG,
min_edge_log=MIN_EDGE_LOG,
edge_log_range=EDGE_LOG_RANGE):
self.filter_tig = FilterTIG()
self.weights_1st_stage = weights_1st_stage
self.filter_tig.update(self.weights_1st_stage)
self.filter_tig.reconstruct(self.weights_1st_stage)
self.scale_space_sizes_idxs = scale_space_sizes_idxs
self.base_log = base_log
self.min_edge_log = min_edge_log
self.edge_log_range = edge_log_range
self.edge = edge
self.num_win_psz = num_win_psz
def predict(self, image, nss=2):
bbs = []
img_h, img_w, nch = image.shape
for size_idx in self.scale_space_sizes_idxs:
w = round(
pow(self.base_log,
size_idx % self.edge_log_range + self.min_edge_log))
h = round(
pow(self.base_log,
size_idx // self.edge_log_range + self.min_edge_log))
if (h > img_h * self.base_log) or (w > img_w * self.base_log):
continue
h = min(h, img_h)
w = min(w, img_w)
new_w = int(round(float(self.edge) * img_w / w))
new_h = int(round(float(self.edge) * img_h / h))
img_resized = cv2.resize(image, (new_w, new_h))
grad = rgb_gradient(img_resized)
match_map = self.filter_tig.match_template(grad)
points = self.filter_tig.non_maxima_suppression(
match_map, nss, self.num_win_psz, False)
ratio_x = w / self.edge
ratio_y = h / self.edge
i_max = min(len(points), self.num_win_psz)
for i in xrange(i_max):
point, score = points[i]
x0 = int(round(point[0] * ratio_x))
y0 = int(round(point[1] * ratio_y))
x1 = min(img_w, int(x0 + w))
y1 = min(img_h, int(y0 + h))
x0 = x0 + 1
y0 = y0 + 1
bbs.append(((x0, y0, x1, y1), score, size_idx))
return bbs
class FirstStagePrediction(object):
def __init__(
self,
weights_1st_stage,
scale_space_sizes_idxs,
num_win_psz=130,
edge=EDGE,
base_log=BASE_LOG,
min_edge_log=MIN_EDGE_LOG,
edge_log_range=EDGE_LOG_RANGE,
):
self.filter_tig = FilterTIG()
self.weights_1st_stage = weights_1st_stage
self.filter_tig.update(self.weights_1st_stage)
self.filter_tig.reconstruct(self.weights_1st_stage)
self.scale_space_sizes_idxs = scale_space_sizes_idxs
self.base_log = base_log
self.min_edge_log = min_edge_log
self.edge_log_range = edge_log_range
self.edge = edge
self.num_win_psz = num_win_psz
def predict(self, image, nss=2):
bbs = []
img_h, img_w, nch = image.shape
for size_idx in self.scale_space_sizes_idxs:
w = round(pow(self.base_log, size_idx % self.edge_log_range + self.min_edge_log))
h = round(pow(self.base_log, size_idx // self.edge_log_range + self.min_edge_log))
if (h > img_h * self.base_log) or (w > img_w * self.base_log):
continue
h = min(h, img_h)
w = min(w, img_w)
new_w = int(round(float(self.edge) * img_w / w))
new_h = int(round(float(self.edge) * img_h / h))
img_resized = cv2.resize(image, (new_w, new_h))
grad = rgb_gradient(img_resized)
match_map = self.filter_tig.match_template(grad)
points = self.filter_tig.non_maxima_suppression(match_map, nss, self.num_win_psz, False)
ratio_x = w / self.edge
ratio_y = h / self.edge
i_max = min(len(points), self.num_win_psz)
for i in xrange(i_max):
point, score = points[i]
x0 = int(round(point[0] * ratio_x))
y0 = int(round(point[1] * ratio_y))
x1 = min(img_w, int(x0 + w))
y1 = min(img_h, int(y0 + h))
x0 = x0 + 1
y0 = y0 + 1
bbs.append(((x0, y0, x1, y1), score, size_idx))
return bbs
