def __init__(self, batch_label, feature_maps):
self.batch_size = batch_label.shape[0]
self.num_predict_features = 6#一共有6个FeatureMap
self.read_dataset = ReadDataset()
self.iou_threshold = IOU_THRESHOLD
self.default_boxes = DefaultBoxes(feature_maps)#从feeture Map中提取出Default box
self.boxes = batch_label
self.predict_boxes = self.default_boxes.generate_default_boxes()
def __init__(self, batch_data, output_features):
self.batch_data = batch_data
self.batch_size = batch_data.shape[0]
self.num_predict_features = 6
self.read_dataset = ReadDataset()
self.default_boxes = DefaultBoxes(feature_map_list=output_features)
self.images, self.boxes = self.read_dataset.read(self.batch_data)
self.predict_boxes = self.default_boxes.generate_default_boxes()
def __offsets_to_true_coordinates(self, pred_boxes, ssd_output):
pred_classes = tf.reshape(tensor=pred_boxes[..., :self.num_classes],
shape=(-1, self.num_classes))
pred_coords = tf.reshape(tensor=pred_boxes[..., self.num_classes:],
shape=(-1, 4))
default_boxes = DefaultBoxes(
feature_map_list=ssd_output).generate_default_boxes()
d_cx, d_cy, d_w, d_h = default_boxes[:, 0:
1], default_boxes[:, 1:
2], default_boxes[:,
2:
3], default_boxes[:,
3:
4]
offset_cx, offset_cy, offset_w, offset_h = pred_coords[:, 0:
1], pred_coords[:,
1:
2], pred_coords[:,
2:
3], pred_coords[:,
3:
4]
true_cx = offset_cx * d_w + d_cx
true_cy = offset_cy * d_h + d_cy
true_w = tf.math.exp(offset_w) * d_w
true_h = tf.math.exp(offset_h) * d_h
true_coords = tf.concat(values=[true_cx, true_cy, true_w, true_h],
axis=-1)
true_classes_and_coords = tf.concat(values=[pred_classes, true_coords],
axis=-1)
true_classes_and_coords = tf.expand_dims(input=true_classes_and_coords,
axis=0)
return true_classes_and_coords
class MakeGT(object):
def __init__(self, batch_data, output_features):
self.batch_data = batch_data
self.batch_size = batch_data.shape[0]
self.num_predict_features = 6
self.read_dataset = ReadDataset()
self.default_boxes = DefaultBoxes(feature_map_list=output_features)
self.images, self.boxes = self.read_dataset.read(self.batch_data)
self.predict_boxes = self.default_boxes.generate_default_boxes()
def ___transform_true_boxes(self):
boxes_xywhc = self.__to_xywhc(self.boxes)
true_boxes_x = boxes_xywhc[..., 0] / IMAGE_WIDTH
true_boxes_y = boxes_xywhc[..., 1] / IMAGE_HEIGHT
true_boxes_w = boxes_xywhc[..., 2] / IMAGE_WIDTH
true_boxes_h = boxes_xywhc[..., 3] / IMAGE_HEIGHT
true_boxes_c = boxes_xywhc[..., -1]
true_boxes = np.stack((true_boxes_x, true_boxes_y, true_boxes_w,
true_boxes_h, true_boxes_c),
axis=-1)
return true_boxes
@staticmethod
def __to_xywhc(boxes):
xy = 0.5 * (boxes[..., 0:2] + boxes[..., 2:4])
wh = boxes[..., 2:4] - boxes[..., 0:2]
c = boxes[..., -1:]
return np.concatenate((xy, wh, c),
axis=-1) # (center_x, center_y, w, h, c)
@staticmethod
def __get_valid_boxes(boxes):
num_of_boxes = boxes.shape[0]
valid_boxes = []
for i in range(num_of_boxes):
if boxes[i, -1] > 0.0:
valid_boxes.append(boxes[i])
valid_boxes = np.array(valid_boxes, dtype=np.float32)
return valid_boxes
def __max_iou(self, box_true, box_pred):
iou_outside = []
for i in range(box_true.shape[0]):
iou_inside = []
for j in range(box_pred.shape[0]):
iou = IOU(box_1=box_true[i], box_2=box_pred[j]).calculate_iou()
iou_inside.append(iou)
iou_outside.append(iou_inside)
iou_array = np.array(iou_outside, dtype=np.float32)
pos_index = np.argmax(iou_array, axis=1)
return pos_index
def label_positive_and_negative_predicted_boxes(self):
true_boxes = self.___transform_true_boxes()
for n in range(self.batch_size):
# shape : (N, 5), where N is the number of valid true boxes for each input image.
valid_true_boxes = self.__get_valid_boxes(true_boxes[n])
true_boxes_for_iou = valid_true_boxes[..., :4]
pos_index = self.__max_iou(true_boxes_for_iou, self.predict_boxes)
def __init__(self, num_classes, overlap_thresh, neg_pos):
self.default_boxes = tf.convert_to_tensor(
DefaultBoxes().generate_boxes()) # Tensor, shape: (先验框总数(8732), 4)
self.num_classes = num_classes
self.threshold = overlap_thresh
self.variance = VARIANCE
self.negpos_ratio = neg_pos
def __init__(self, model, num_classes):
self.model = model
self.priors = DefaultBoxes().generate_boxes(
) # (num_priors,4) num_priors表示anchor总数，在SSD300中为8732
self.top_k = MAX_BOXES_NUM
self.num_classes = num_classes
self.variance = VARIANCE
self.conf_thresh = CONFIDENCE_THRESHOLD
self.nms_thresh = NMS_THRESHOLD
class MakeGT(object):
def __init__(self, batch_label, feature_maps):
self.batch_size = batch_label.shape[0]
self.num_predict_features = 6#一共有6个FeatureMap
self.read_dataset = ReadDataset()
self.iou_threshold = IOU_THRESHOLD
self.default_boxes = DefaultBoxes(feature_maps)#从feeture Map中提取出Default box
self.boxes = batch_label
self.predict_boxes = self.default_boxes.generate_default_boxes()
def ___transform_true_boxes(self):#将ground Truth的Lable的位置转换成Tensor
boxes_xywhc = self.__to_xywhc(self.boxes)
true_boxes_x = boxes_xywhc[..., 0] / IMAGE_WIDTH
true_boxes_y = boxes_xywhc[..., 1] / IMAGE_HEIGHT
true_boxes_w = boxes_xywhc[..., 2] / IMAGE_WIDTH
true_boxes_h = boxes_xywhc[..., 3] / IMAGE_HEIGHT
true_boxes_c = boxes_xywhc[..., -1]
true_boxes = np.stack((true_boxes_x, true_boxes_y, true_boxes_w, true_boxes_h, true_boxes_c), axis=-1)
return true_boxes
@staticmethod
def __to_xywhc(boxes):
xy = 0.5 * (boxes[..., 0:2] + boxes[..., 2:4])
wh = boxes[..., 2:4] - boxes[..., 0:2]
c = boxes[..., -1:]
return np.concatenate((xy, wh, c), axis=-1) # (center_x, center_y, w, h, c)
@staticmethod
def __get_valid_boxes(boxes):#计算出有效的box，也就是Box的各个参数都能在feature Map中找到
num_of_boxes = boxes.shape[0]
valid_boxes = []
for i in range(num_of_boxes):
if boxes[i, -1] > 0.0:
valid_boxes.append(boxes[i])
valid_boxes = np.array(valid_boxes, dtype=np.float32)
return valid_boxes
def __label_positive_and_negative_predicted_boxes(self, box_true, box_pred):#通过IoU的计算，来其确定pred中有多少是Positiv的
box_true_coord = box_true[..., :4]#box的[cx，cy，w，h]
box_true_class = box_true[..., -1]#class Lable
box_pred_assigned = np.zeros_like(box_pred, dtype=np.float32)
iou_outside = []
for i in range(box_true_coord.shape[0]):
iou_inside = []
for j in range(box_pred.shape[0]):
iou = IOU(box_1=box_true_coord[i], box_2=box_pred[j]).calculate_iou()
iou_inside.append(iou)
iou_outside.append(iou_inside)
iou_array = np.array(iou_outside, dtype=np.float32) # shape: (num_of_true_boxes, total_num_of_default_boxes)
iou_max = np.max(iou_array, axis=0)#计算出各个列中，最大的IoU值
max_index = np.argmax(iou_array, axis=0)#找出最大值对应的Index，在哪一列
max_index_class = np.zeros_like(max_index, dtype=np.float32)#生成形状类似的全0 数组
for k in range(max_index.shape[0]):
max_index_class[k] = box_true_class[max_index[k]]
box_pred_assigned[k] = self.__get_offset(box_true=box_true_coord[max_index[k]], box_pred=box_pred[k])
pos_boolean = np.where(iou_max > self.iou_threshold, 1.0, 0.0) # 1 for positive, 0 for negative
pos_class_index = max_index_class * pos_boolean
pos_class_index = pos_class_index.reshape((-1, 1))
labeled_box_pred = np.concatenate((box_pred_assigned, pos_class_index), axis=-1)
return labeled_box_pred
@staticmethod
def __get_offset(box_true, box_pred):
d_cx, d_cy, d_w, d_h = box_pred
g_cx, g_cy, g_w, g_h = box_true
g_cx = (g_cx - d_cx) / d_w
g_cy = (g_cy - d_cy) / d_h
g_w = np.log(g_w / d_w)
g_h = np.log(g_h / d_h)
return np.stack([g_cx, g_cy, g_w, g_h], axis=0)
def generate_gt_boxes(self):
true_boxes = self.___transform_true_boxes() # shape: (batch_size, MAX_BOXES_PER_IMAGE, 5)
gt_boxes_list = []
for n in range(self.batch_size):
# shape : (N, 5), where N is the number of valid true boxes for each input image.
valid_true_boxes = self.__get_valid_boxes(true_boxes[n])
gt_boxes = self.__label_positive_and_negative_predicted_boxes(valid_true_boxes, self.predict_boxes)
gt_boxes_list.append(gt_boxes)
batch_gt_boxes = np.stack(gt_boxes_list, axis=0) # shape: (batch_size, total_num_of_default_boxes, 5)
return tf.convert_to_tensor(value=batch_gt_boxes, dtype=tf.dtypes.float32)
class MakeGT(object):
def __init__(self, batch_label, feature_maps):
self.batch_size = batch_label.shape[0]
self.num_predict_features = 6
self.read_dataset = ReadDataset()
self.iou_threshold = IOU_THRESHOLD
self.default_boxes = DefaultBoxes(feature_maps)
self.boxes = batch_label
self.predict_boxes = self.default_boxes.generate_default_boxes()
def ___transform_true_boxes(self):
boxes_xywhc = self.__to_xywhc(self.boxes)
true_boxes_x = boxes_xywhc[..., 0] / IMAGE_WIDTH
true_boxes_y = boxes_xywhc[..., 1] / IMAGE_HEIGHT
true_boxes_w = boxes_xywhc[..., 2] / IMAGE_WIDTH
true_boxes_h = boxes_xywhc[..., 3] / IMAGE_HEIGHT
true_boxes_c = boxes_xywhc[..., -1]
true_boxes = np.stack((true_boxes_x, true_boxes_y, true_boxes_w,
true_boxes_h, true_boxes_c),
axis=-1)
return true_boxes
@staticmethod
def __to_xywhc(boxes):
xy = 0.5 * (boxes[..., 0:2] + boxes[..., 2:4])
wh = boxes[..., 2:4] - boxes[..., 0:2]
c = boxes[..., -1:]
return np.concatenate((xy, wh, c),
axis=-1) # (center_x, center_y, w, h, c)
@staticmethod
def __get_valid_boxes(boxes):
num_of_boxes = boxes.shape[0]
valid_boxes = []
for i in range(num_of_boxes):
if boxes[i, -1] > 0.0:
valid_boxes.append(boxes[i])
valid_boxes = np.array(valid_boxes, dtype=np.float32)
return valid_boxes
def __label_positive_and_negative_predicted_boxes(self, box_true,
box_pred):
box_true_coord = box_true[..., :4]
box_true_class = box_true[..., -1]
box_pred_assigned = np.zeros_like(box_pred, dtype=np.float32)
iou_outside = []
for i in range(box_true_coord.shape[0]):
iou_inside = []
for j in range(box_pred.shape[0]):
iou = IOU(box_1=box_true_coord[i],
box_2=box_pred[j]).calculate_iou()
iou_inside.append(iou)
iou_outside.append(iou_inside)
iou_array = np.array(
iou_outside, dtype=np.float32
) # shape: (num_of_true_boxes, total_num_of_default_boxes)
iou_max = np.max(iou_array, axis=0)
max_index = np.argmax(iou_array, axis=0)
max_index_class = np.zeros_like(max_index, dtype=np.float32)
for k in range(max_index.shape[0]):
max_index_class[k] = box_true_class[max_index[k]]
box_pred_assigned[k] = box_true_coord[max_index[k]]
pos_boolean = np.where(iou_max > self.iou_threshold, 1,
0) # 1 for positive, 0 for negative
pos_class_index = max_index_class * pos_boolean
pos_class_index = pos_class_index.reshape((-1, 1))
labeled_box_pred = np.concatenate((box_pred_assigned, pos_class_index),
axis=-1)
return labeled_box_pred
def generate_gt_boxes(self):
true_boxes = self.___transform_true_boxes(
) # shape: (batch_size, MAX_BOXES_PER_IMAGE, 5)
gt_boxes_list = []
for n in range(self.batch_size):
# shape : (N, 5), where N is the number of valid true boxes for each input image.
valid_true_boxes = self.__get_valid_boxes(true_boxes[n])
gt_boxes = self.__label_positive_and_negative_predicted_boxes(
valid_true_boxes, self.predict_boxes)
gt_boxes_list.append(gt_boxes)
batch_gt_boxes = np.stack(
gt_boxes_list,
axis=0) # shape: (batch_size, total_num_of_default_boxes, 5)
return tf.convert_to_tensor(value=batch_gt_boxes,
dtype=tf.dtypes.float32)
