def get_zoom_negtive_samples(self):
# feature map size
n, h, w, ratio = len(self.masks), self.feat_h, self.feat_w, self.ratio
match = np.ones((h * w, n), np.int)
win_bbs = self.win_bbs
objn_win_cens = self.objn_win_cens
win_pts = self.win_pts
'''
# condition 1: not too large or too small
for i in range(n):
match[:, i] = (self.bbs_hw[i].max() >= SLIDING_WINDOW_SIZE * ratio * OBJN_LOWER_BOUND_RATIO).all() & (self.bbs_hw[i].max() <= SLIDING_WINDOW_SIZE * ratio * OBJN_UPPER_BOUND_RATIO).all()
'''
# condition 2: roughly contained
for i in range(n):
match[:, i] = match[:, i] & pts_in_bbs(self.centers[i], win_bbs)
# condition 3: roughly centered
for i in range(n):
match[:,
i] = match[:, i] & pts_in_bbs(self.centers[i], objn_win_cens)
# choose the closest one
dist = match * -1e9
for i in range(n):
dist[:, i] += np.linalg.norm(win_pts - self.centers[i], axis=1)
obj_ids = np.argmin(dist, axis=1)
match[np.arange(h * w), obj_ids] += 1
match[match < 2] = 0
match[match == 2] = 1
self.zoom_negtive_match = match
secondary_objns = np.zeros((h * w))
secondary_objns[match.any(axis=1)] = 1
objns = np.zeros((h * w))
objns[self.objn_match.any(axis=1)] = 1
negtive_samples = np.zeros((h * w))
negtive_samples[(secondary_objns == 1) & (objns == 0)] = 1
if self.zoom_negtive_samples is None:
self.zoom_negtive_samples = negtive_samples
else:
self.zoom_negtive_samples = np.concatenate(
(self.zoom_negtive_samples, negtive_samples), axis=0)
def get_zoom_negtive_samples(self):
n, h, w, ratio = len(self.masks), self.feat_h, self.feat_w, self.ratio
match = np.ones((h * w, n), np.int8)
win_bbs = self.win_bbs
objn_win_cens = self.objn_win_cens
win_pts = self.win_pts
# condition 2: roughly contained
for i in range(n):
match[:, i] = match[:, i] & pts_in_bbs(self.centers[i], win_bbs)
# condition 3: roughly centered
for i in range(n):
match[:,
i] = match[:, i] & pts_in_bbs(self.centers[i], objn_win_cens)
# choose the closest one
dist = match * -1e9
for i in range(n):
dist[:, i] += np.linalg.norm(win_pts - self.centers[i], axis=1)
obj_ids = np.argmin(dist, axis=1)
match[np.arange(h * w), obj_ids] += 1
match[match < 2] = 0
match[match == 2] = 1
self.zoom_negtive_match = match
secondary_objns = np.zeros((h * w))
secondary_objns[match.any(axis=1)] = 1
objns = np.zeros((h * w))
objns[self.objn_match.any(axis=1)] = 1
negtive_samples = np.zeros((h * w))
negtive_samples[(secondary_objns == 1) & (objns == 0)] = 1
if self.zoom_negtive_samples is None:
self.zoom_negtive_samples = negtive_samples
else:
self.zoom_negtive_samples = np.concatenate(
(self.zoom_negtive_samples, negtive_samples), axis=0)
def find_matched_masks(self):
n, h, w, ratio = len(self.masks), self.feat_h, self.feat_w, self.ratio
win_pts = np.array((np.arange(h * w, dtype=np.int) / w,
np.arange(h * w, dtype=np.int) % w))
win_pts = win_pts.transpose((1, 0)).astype(np.float)
win_pts *= ratio
self.win_pts = win_pts
objn_win_cens = np.hstack(
(win_pts - (SLIDING_WINDOW_SIZE * ratio * OBJN_CENTER_RATIO / 2.0),
win_pts +
(SLIDING_WINDOW_SIZE * ratio * OBJN_CENTER_RATIO / 2.0)))
mask_win_cens = np.hstack(
(win_pts - (SLIDING_WINDOW_SIZE * ratio * MASK_CENTER_RATIO / 2.0),
win_pts +
(SLIDING_WINDOW_SIZE * ratio * MASK_CENTER_RATIO / 2.0)))
self.objn_win_cens = objn_win_cens
self.mask_win_cens = mask_win_cens
win_bbs = np.hstack((win_pts - (SLIDING_WINDOW_SIZE * ratio / 2.0),
win_pts + (SLIDING_WINDOW_SIZE * ratio / 2.0)))
self.win_bbs = win_bbs
self.objn_match = np.ones((h * w, n), np.int8)
self.mask_match = np.ones((h * w, n), np.int8)
# condition 1: neither too large nor too small
for i in range(n):
self.objn_match[:, i] = (
self.bbs_hw[i].max() >=
SLIDING_WINDOW_SIZE * ratio * OBJN_LOWER_BOUND_RATIO).all() & (
self.bbs_hw[i].max() <= SLIDING_WINDOW_SIZE * ratio *
OBJN_UPPER_BOUND_RATIO).all()
for i in range(n):
self.mask_match[:, i] = (
self.bbs_hw[i].max() >=
SLIDING_WINDOW_SIZE * ratio * MASK_LOWER_BOUND_RATIO).all() & (
self.bbs_hw[i].max() <= SLIDING_WINDOW_SIZE * ratio *
MASK_UPPER_BOUND_RATIO).all()
# condition 2: roughly contained
for i in range(n):
self.objn_match[:, i] = self.objn_match[:, i] & pts_in_bbs(
self.centers[i], win_bbs)
for i in range(n):
self.mask_match[:, i] = self.mask_match[:, i] & pts_in_bbs(
self.centers[i], win_bbs)
# condition 3: roughly centered
for i in range(n):
self.objn_match[:, i] = self.objn_match[:, i] & pts_in_bbs(
self.centers[i], objn_win_cens)
for i in range(n):
self.mask_match[:, i] = self.mask_match[:, i] & pts_in_bbs(
self.centers[i], mask_win_cens)
# choose the closest one
dist = self.objn_match * -1e9
for i in range(n):
dist[:, i] += np.linalg.norm(win_pts - self.centers[i], axis=1)
obj_ids = np.argmin(dist, axis=1)
self.objn_match[np.arange(h * w), obj_ids] += 1
self.objn_match[self.objn_match < 2] = 0
self.objn_match[self.objn_match == 2] = 1
dist = self.mask_match * -1e9
for i in range(n):
dist[:, i] += np.linalg.norm(win_pts - self.centers[i], axis=1)
obj_ids = np.argmin(dist, axis=1)
self.mask_match[np.arange(h * w), obj_ids] += 1
self.mask_match[self.mask_match < 2] = 0
self.mask_match[self.mask_match == 2] = 1
self.get_zoom_negtive_samples()
self.get_shift_negtive_samples()
def find_matched_masks(self):
# feature map size
n, h, w, ratio = len(self.masks), self.feat_h, self.feat_w, self.ratio
# win_pts: [h * w, 2]
win_pts = np.array((np.arange(h * w, dtype=np.int) / w,
np.arange(h * w, dtype=np.int) % w))
win_pts = win_pts.transpose((1, 0)).astype(np.float)
win_pts *= ratio
self.win_pts = win_pts
# objn_win_cens: [h * w, 4]
# mask_win_cens: [h * w, 4]
objn_win_cens = np.hstack(
(win_pts - (SLIDING_WINDOW_SIZE * ratio * OBJN_CENTER_RATIO / 2.0),
win_pts +
(SLIDING_WINDOW_SIZE * ratio * OBJN_CENTER_RATIO / 2.0)))
mask_win_cens = np.hstack(
(win_pts - (SLIDING_WINDOW_SIZE * ratio * MASK_CENTER_RATIO / 2.0),
win_pts +
(SLIDING_WINDOW_SIZE * ratio * MASK_CENTER_RATIO / 2.0)))
self.objn_win_cens = objn_win_cens
self.mask_win_cens = mask_win_cens
# origin image
img = self.image.copy()
img += RGB_MEAN
# visualize centers of windows
'''
centers = win_cens + SLIDING_WINDOW_SIZE * CENTER_RATIO * ratio / 2
bbs = centers[:20, :]
visualize_bbs(img, bbs)
'''
# win_bbs: [h * w, 4]
win_bbs = np.hstack((win_pts - (SLIDING_WINDOW_SIZE * ratio / 2.0),
win_pts + (SLIDING_WINDOW_SIZE * ratio / 2.0)))
self.win_bbs = win_bbs
# [h * w, label_num)
self.objn_match = np.ones((h * w, n), np.int)
self.mask_match = np.ones((h * w, n), np.int)
# condition 1: neither too large nor too small
for i in range(n):
self.objn_match[:, i] = (
self.bbs_hw[i].max() >=
SLIDING_WINDOW_SIZE * ratio * OBJN_LOWER_BOUND_RATIO).all() & (
self.bbs_hw[i].max() <= SLIDING_WINDOW_SIZE * ratio *
OBJN_UPPER_BOUND_RATIO).all()
for i in range(n):
self.mask_match[:, i] = (
self.bbs_hw[i].max() >=
SLIDING_WINDOW_SIZE * ratio * MASK_LOWER_BOUND_RATIO).all() & (
self.bbs_hw[i].max() <= SLIDING_WINDOW_SIZE * ratio *
MASK_UPPER_BOUND_RATIO).all()
# condition 2: roughly contained
for i in range(n):
self.objn_match[:, i] = self.objn_match[:, i] & pts_in_bbs(
self.centers[i], win_bbs)
for i in range(n):
self.mask_match[:, i] = self.mask_match[:, i] & pts_in_bbs(
self.centers[i], win_bbs)
# condition 3: roughly centered
for i in range(n):
self.objn_match[:, i] = self.objn_match[:, i] & pts_in_bbs(
self.centers[i], objn_win_cens)
for i in range(n):
self.mask_match[:, i] = self.mask_match[:, i] & pts_in_bbs(
self.centers[i], mask_win_cens)
# choose the closest one
dist = self.objn_match * -1e9
for i in range(n):
dist[:, i] += np.linalg.norm(win_pts - self.centers[i], axis=1)
obj_ids = np.argmin(dist, axis=1)
self.objn_match[np.arange(h * w), obj_ids] += 1
self.objn_match[self.objn_match < 2] = 0
self.objn_match[self.objn_match == 2] = 1
dist = self.mask_match * -1e9
for i in range(n):
dist[:, i] += np.linalg.norm(win_pts - self.centers[i], axis=1)
obj_ids = np.argmin(dist, axis=1)
self.mask_match[np.arange(h * w), obj_ids] += 1
self.mask_match[self.mask_match < 2] = 0
self.mask_match[self.mask_match == 2] = 1
# visualize matched sliding window
'''
bbs = np.array((np.nonzero(self.objn_match)[0]/w, np.nonzero(self.objn_match)[0]%w)) * ratio
bbs = np.vstack((bbs-SLIDING_WINDOW_SIZE/2*ratio, bbs+SLIDING_WINDOW_SIZE/2*ratio)).transpose((1, 0))
visualize_bbs(img, bbs[:16])
visualize_bbs(img, bbs[16:])
'''
# get hard training samples
self.get_zoom_negtive_samples()
self.get_shift_negtive_samples()