def _restore_rect(self, rect, tensor_lr, tensor_tb):
assert tensor_lr.shape == tensor_tb.shape
assert len(tensor_lr.shape) == 3
assert isinstance(rect, Rect)
map_h, map_w, _ = tensor_lr.shape
sum_l, sum_r = 0., 0.
for i in range(rect.top, rect.bottom + 1):
sum_l += tensor_lr[i][rect.left][0]
sum_r += -tensor_lr[i][rect.right][1]
sum_t, sum_b = 0., 0.
for i in range(rect.left, rect.right + 1):
sum_t += tensor_tb[rect.top][i][0]
sum_b += - tensor_tb[rect.bottom][i][1]
height, width = rect.bottom - rect.top + 1, rect.right - rect.left + 1
ans = Rect(rect.left + sum_l / height, rect.top + sum_t / width,
rect.right + 1 + sum_r / height, rect.bottom + 1 + sum_b / width)
#print ans
ans.stretch(1.0 / map_h, 1.0 / map_w)
#print ans
return ans
def _restore_rects(self,
lr,
tb,
cls,
num_poolings,
threshold=None,
top=None):
# print ratio
map_h, map_w = lr.shape[1:3]
if self.verbose:
print map_h, map_w
if cls is not None:
print cls.shape
left_logit = lr[:, :, :, 0]
right_logit = lr[:, :, :, 1]
top_logit = tb[:, :, :, 0]
bottom_logit = tb[:, :, :, 1]
if cls is not None:
bg_logit = np.exp(cls[:, :, :, 0])
bar_logit = np.exp(cls[:, :, :, 1])
prob = np.exp(bar_logit) / (np.exp(bar_logit) + np.exp(bg_logit))
dleft = left_logit
dright = 1.0 - right_logit
dtop = top_logit
dbottom = 1.0 - bottom_logit
def cut_top(res):
res = sorted(res, reverse=True, key=lambda val: val[0])
if top is not None:
res = res[:top]
return res
res = []
for y in range(0, map_h):
for x in range(0, map_w):
dl, dr, dt, db = dleft[0, y,
x], dright[0, y,
x], dtop[0, y,
x], dbottom[0, y, x]
#dy, dx, sh, sw = 0,0,0,0
#sh, sw = 0, 0
#dy, dx = 0, 0
# print cls.shape
conf = prob[0, y, x] if cls is not None else 0
if self.verbose and conf > 0:
print conf, y, x, dl, dr, dt, db
if threshold is not None or conf > threshold:
rect = Rect(x + dl * 2, y + dt * 2, x + dr * 2, y + db * 2)
rect.stretch(1 << (num_poolings), 1 << (num_poolings))
if self.verbose:
print rect
res.append((conf, rect))
if top is not None and len(res) > 2 * top:
res = cut_top(res)
return cut_top(res)
