def encode(self, gt_data, overlap_threshold=0.5, debug=False):
'''
gt_data의 예시:
image 하나당 [x1,y1, x2,y2, x3,y3, x4,y4, 6_one_hot_class]의 데이터가 box갯수만큼 있음
[[0.56833333 0.20300752 0.97666667 0.33333333 0.97 0.47619048
0.55833333 0.34586466 0. 0. 0. 0.
0. 1. ]
[0.11666667 0.45864662 0.14166667 0.45614035 0.14 0.49874687
0.115 0.50125313 0. 1. 0. 0.
0. 0. ]
[0.88333333 0.47869674 0.91166667 0.48120301 0.91 0.50626566
0.88166667 0.5037594 0. 1. 0. 0.
0. 0. ]
[0.03166667 0.30075188 0.42666667 0.160401 0.44666667 0.30075188
0.03666667 0.46867168 0. 1. 0. 0.
0. 0. ]
[0.15833333 0.39598997 0.42 0.40601504 0.415 0.48120301
0.15333333 0.47368421 0. 1. 0. 0.
0. 0. ]
[0.58166667 0.41854637 0.85833333 0.4235589 0.86 0.4962406
0.58166667 0.49373434 0. 1. 0. 0.
0. 0. ]
[0.63 0.57894737 0.68333333 0.60150376 0.67666667 0.64160401
0.62333333 0.61904762 0. 1. 0. 0.
0. 0. ]
[0.46 0.71428571 0.555 0.71428571 0.555 0.74937343
0.46 0.74937343 0. 0. 0. 0.
0. 1. ]]
위에서 정의한 prior box들의 정보를 가지고 prior box와 gt_data를 match시켜서
shape = [priormap갯수(엄청큼), 4+8+5+confidence_of_class] 인 넘파이배열을 return한다.
confidence class에서는 4+8+5의 정보가 gt_data와의 iou가 0.5를 넘는게 없다면 background class로 분류된다.
'''
'''
array([[0.09446537, 0.64231986, 0.6947241 , 0.80873207, 0.66108891,
1.05911375, 0.05824286, 0.92018245, 0. , 0. ,
0. , 0. , 0. , 1. ]]) error찾기
'''
# image에 box가 없을경우
if gt_data.shape[0] == 0:
print('gt_data', type(gt_data), gt_data.shape)
eps = 1e-15 #수정 한상준13
num_classes = 6
num_priors = self.priors.shape[0]
gt_polygons = np.copy(gt_data[:, :8]) # normalized quadrilaterals
gt_rboxes = np.array(
[polygon_to_rbox3(np.reshape(p, (-1, 2))) for p in gt_data[:, :8]])
# minimum horizontal bounding rectangles
gt_xmin = np.min(gt_data[:, 0:8:2], axis=1) #shape : [nb_boxes]
gt_ymin = np.min(gt_data[:, 1:8:2], axis=1)
gt_xmax = np.max(gt_data[:, 0:8:2], axis=1)
gt_ymax = np.max(gt_data[:, 1:8:2], axis=1)
gt_boxes = self.gt_boxes = np.array(
[gt_xmin, gt_ymin, gt_xmax, gt_ymax]).T # shape : [nb_boxes, 4]
# normalized xmin, ymin, xmax, ymax
gt_one_hot = gt_data[:, 8:] # shape : [nb_boxes, 6]
gt_iou = np.array([iou(b, self.priors_norm) for b in gt_boxes]).T
# shape of self.priors_norm : [Nums_priors, min_xy + max_xy+ variance]
# b = [4], self.priors_norm = [Nums_priors, 4+4]
# shape of gt_iou : [[Nums_priors] for b in gt_boxes].transpose = [nb_boxes, Nums_priors].transpose = [Num_priors, nb_boxes]
# assigne gt to priors
max_idxs = np.argmax(gt_iou, axis=1) # shape : [Num_priors]
max_val = gt_iou[np.arange(num_priors),
max_idxs] # shape: [Num_priors]
prior_mask = max_val > overlap_threshold #IOU값의 maximum이 0.5가 넘는 prior들만 True값을 줌
match_indices = max_idxs[
prior_mask] # shape : [Num_priors - False_priors]
self.match_indices = dict(
zip(list(np.ix_(prior_mask)[0]), list(match_indices)))
# {prior_1 : max_idx, prior_3 : max_idx, prior_4 : max_idx, prior_7 : max_idx, ...}
# prior labels
confidence = np.zeros((num_priors, num_classes))
confidence[:, 0] = 1 # 일단 모든 prior들을 background 로 정의
confidence[prior_mask] = gt_one_hot[
match_indices] # mask가 True인것(iou>0.5)만 one_hot_class를 새로 매겨줌
gt_xy = (gt_boxes[:, 2:4] +
gt_boxes[:, 0:2]) / 2. #shape : [nb_boxes,2]
gt_wh = gt_boxes[:, 2:4] - gt_boxes[:, 0:2] #shape : [nb_boxes,2]
gt_xy = gt_xy[
match_indices] #shape : [True_priors,2] , True_priors = Num_priors - False_priors
gt_wh = gt_wh[match_indices] #shape : [True_priors,2]
gt_polygons = gt_polygons[match_indices] #shape : [True_priors,8]
gt_rboxes = gt_rboxes[match_indices] #shape : [True_priors,5]
priors_xy = self.priors_xy[
prior_mask] / self.image_size # = self.priors_xy_norm[prior_mask]
priors_wh = self.priors_wh[
prior_mask] / self.image_size # = self.priors_wh_norm[prior_mask]
variances_xy = self.priors_variances[prior_mask, 0:2]
variances_wh = self.priors_variances[prior_mask, 2:4]
# compute local offsets for
# gt_x = prior_x + prior_w * x_label
# x_label = x_label / 0.1
# gt_w = exp(w_label) * prior_w
# w_label = w_label / 0.2
offsets = np.zeros((num_priors, 4))
offsets[prior_mask, 0:2] = (gt_xy - priors_xy) / priors_wh
offsets[prior_mask, 2:4] = np.log(gt_wh / priors_wh) #수정 한상준13
offsets[prior_mask, 0:2] /= variances_xy
offsets[prior_mask, 2:4] /= variances_wh
# compute local offsets for quadrilaterals
# gt_x1 = prior_x1 + prior_w * x1_label
# x1_label = x1_label / 0.1
offsets_quads = np.zeros((num_priors, 8))
priors_xy_minmax = np.hstack(
[priors_xy - priors_wh / 2, priors_xy + priors_wh / 2])
ref = priors_xy_minmax[:, (0, 1, 2, 1, 2, 3, 0, 3)] # corner points
offsets_quads[prior_mask, :] = (gt_polygons - ref) / np.tile(
priors_wh, (1, 4)) / np.tile(variances_xy, (1, 4))
# compute local offsets for rotated bounding boxes
# gt_x1 = prior_x1 + prior_w * x1_label
# x1_label = x1_label / 0.1
# gt_h = exp(h_label) * prior_h
# h_label = h_label / 0.2
offsets_rboxs = np.zeros((num_priors, 5))
offsets_rboxs[prior_mask, 0:2] = (gt_rboxes[:, 0:2] -
priors_xy) / priors_wh / variances_xy
offsets_rboxs[prior_mask, 2:4] = (gt_rboxes[:, 2:4] -
priors_xy) / priors_wh / variances_xy
offsets_rboxs[prior_mask, 4] = np.log(
gt_rboxes[:, 4] / priors_wh[:, 1]) / variances_wh[:, 1] #수정 한상준13
return np.concatenate(
[offsets, offsets_quads, offsets_rboxs, confidence], axis=1)
def encode(self, gt_data, overlap_threshold=0.5, debug=False):
# calculation is done with normalized sizes
# TODO: empty ground truth
if gt_data.shape[0] == 0:
print('gt_data', type(gt_data), gt_data.shape)
num_classes = 2
num_priors = self.priors.shape[0]
gt_polygons = np.copy(gt_data[:, :8]) # normalized quadrilaterals
gt_rboxes = np.array(
[polygon_to_rbox3(np.reshape(p, (-1, 2))) for p in gt_data[:, :8]])
# minimum horizontal bounding rectangles
gt_xmin = np.min(gt_data[:, 0:8:2], axis=1)
gt_ymin = np.min(gt_data[:, 1:8:2], axis=1)
gt_xmax = np.max(gt_data[:, 0:8:2], axis=1)
gt_ymax = np.max(gt_data[:, 1:8:2], axis=1)
gt_boxes = self.gt_boxes = np.array(
[gt_xmin, gt_ymin, gt_xmax,
gt_ymax]).T # normalized xmin, ymin, xmax, ymax
gt_class_idx = np.asarray(gt_data[:, -1] + 0.5, dtype=np.int)
gt_one_hot = np.zeros([len(gt_class_idx), num_classes])
gt_one_hot[range(len(gt_one_hot)),
gt_class_idx] = 1 # one_hot classes including background
gt_iou = np.array([iou(b, self.priors_norm) for b in gt_boxes]).T
# assigne gt to priors
max_idxs = np.argmax(gt_iou, axis=1)
max_val = gt_iou[np.arange(num_priors), max_idxs]
prior_mask = max_val > overlap_threshold
match_indices = max_idxs[prior_mask]
self.match_indices = dict(
zip(list(np.ix_(prior_mask)[0]), list(match_indices)))
# prior labels
confidence = np.zeros((num_priors, num_classes))
confidence[:, 0] = 1
confidence[prior_mask] = gt_one_hot[match_indices]
gt_xy = (gt_boxes[:, 2:4] + gt_boxes[:, 0:2]) / 2.
gt_wh = gt_boxes[:, 2:4] - gt_boxes[:, 0:2]
gt_xy = gt_xy[match_indices]
gt_wh = gt_wh[match_indices]
gt_polygons = gt_polygons[match_indices]
gt_rboxes = gt_rboxes[match_indices]
priors_xy = self.priors_xy[prior_mask] / self.image_size
priors_wh = self.priors_wh[prior_mask] / self.image_size
variances_xy = self.priors_variances[prior_mask, 0:2]
variances_wh = self.priors_variances[prior_mask, 2:4]
# compute local offsets for
offsets = np.zeros((num_priors, 4))
offsets[prior_mask, 0:2] = (gt_xy - priors_xy) / priors_wh
offsets[prior_mask, 2:4] = np.log(gt_wh / priors_wh)
offsets[prior_mask, 0:2] /= variances_xy
offsets[prior_mask, 2:4] /= variances_wh
# compute local offsets for quadrilaterals
offsets_quads = np.zeros((num_priors, 8))
priors_xy_minmax = np.hstack(
[priors_xy - priors_wh / 2, priors_xy + priors_wh / 2])
#ref = np.tile(priors_xy, (1,4))
ref = priors_xy_minmax[:, (0, 1, 2, 1, 2, 3, 0, 3)] # corner points
offsets_quads[prior_mask, :] = (gt_polygons - ref) / np.tile(
priors_wh, (1, 4)) / np.tile(variances_xy, (1, 4))
# compute local offsets for rotated bounding boxes
offsets_rboxs = np.zeros((num_priors, 5))
offsets_rboxs[prior_mask, 0:2] = (gt_rboxes[:, 0:2] -
priors_xy) / priors_wh / variances_xy
offsets_rboxs[prior_mask, 2:4] = (gt_rboxes[:, 2:4] -
priors_xy) / priors_wh / variances_xy
offsets_rboxs[prior_mask, 4] = np.log(
gt_rboxes[:, 4] / priors_wh[:, 1]) / variances_wh[:, 1]
return np.concatenate(
[offsets, offsets_quads, offsets_rboxs, confidence], axis=1)