tensorflow
FLAGS.train_image_size = 300
out_shape = [FLAGS.train_image_size] * 2
anchor_creator = anchor_manipulator.AnchorCreator(out_shape,
layers_shapes = ,
anchor_scales = [(0.1,), (0.2,), (0.375,), (0.55,), (0.725,), (0.9,)],
extra_anchor_scales = ,
anchor_ratios = ,
layer_steps = )
class AnchorCreator(object):
def __init__(self, img_shape, layers_shapes, anchor_scales, extra_anchor_scales, anchor_ratios, layer_steps):
super(AnchorCreator, self).__init__()
# img_shape -> (height, width)
self._img_shape = img_shape
self._layers_shapes = layers_shapes
self._anchor_scales = anchor_scales
self._extra_anchor_scales = extra_anchor_scales
self._anchor_ratios = anchor_ratios
self._layer_steps = layer_steps
self._anchor_offset = [0.5] * len(self._layers_shapes)
【入参】：
# img_shape: 原始图形尺寸
self._img_shape = [300, 300]
# layers_shapes：输出层特征图尺寸大小
self._layers_shapes = [(38, 38), (19, 19), (10, 10), (5, 5), (3, 3), (1, 1)]
# anchor_scales: 各层锚框的尺度缩小因子
self._anchor_scales = [(0.1,), (0.2,), (0.375,), (0.55,), (0.725,), (0.9,)]
# extra_anchor_scales: 每层方框的额外尺度缩小因子
self._extra_anchor_scales = [(0.1414,), (0.2739,), (0.4541,), (0.6315,), (0.8078,), (0.9836,)]
# anchor_ratios: 每层锚框的宽高比
self._anchor_ratios = [(1., 2., .5), (1., 2., 3., .5, 0.3333), (1., 2., 3., .5, 0.3333), (1., 2., 3., .5, 0.3333), (1., 2., .5), (1., 2., .5)]
# layer_steps：原图与输出层特征图的尺寸比例
self._layer_steps = [8, 16, 32, 64, 100, 300]
# anchor_offset：锚框的起点偏移
self._anchor_offset = [0.5]*6
【出参】:
all_anchors
all_num_anchors_depth
all_num_anchors_spatial
出参初始化
all_anchors = []
all_num_anchors_depth = []
all_num_anchors_spatial = []
【计算过程】：
for layer_index, layer_shape in enumerate(self._layers_shapes):
anchors_this_layer = self.get_layer_anchors(layer_shape,
self._anchor_scales[layer_index],
self._extra_anchor_scales[layer_index],
self._anchor_ratios[layer_index],
self._layer_steps[layer_index],
self._anchor_offset[layer_index])
all_anchors.append(anchors_this_layer[:-2])
all_num_anchors_depth.append(anchors_this_layer[-2])
all_num_anchors_spatial.append(anchors_this_layer[-1])
以第一层为例：
[参数列表]
layer_shape = [38,38]
layer_index = 0
anchor_scale = (0.1,)
extra_anchor_scale = (0.1414,)
anchor_ratio = (1., 2., .5)
layer_step = 8
offset = 0.5
def get_layer_anchors(self, layer_shape, anchor_scale, extra_anchor_scale, anchor_ratio, layer_step, offset = 0.5):
with tf.name_scope('get_layer_anchors'):
# x_on_layer, y_on_layer = tf.meshgrid(tf.range(38), tf.range(38))
x_on_layer, y_on_layer = tf.meshgrid(tf.range(layer_shape[1]), tf.range(layer_shape[0]))
y_on_image = (tf.cast(y_on_layer, tf.float32) + offset) * layer_step / self._img_shape[0]
x_on_image = (tf.cast(x_on_layer, tf.float32) + offset) * layer_step / self._img_shape[1]
num_anchors_along_depth = len(anchor_scale) * len(anchor_ratio) + len(extra_anchor_scale)
num_anchors_along_spatial = layer_shape[1] * layer_shape[0]
list_h_on_image = []
list_w_on_image = []
global_index = 0
# for square anchors
for _, scale in enumerate(extra_anchor_scale):
list_h_on_image.append(scale)
list_w_on_image.append(scale)
global_index += 1
# for other aspect ratio anchors
for scale_index, scale in enumerate(anchor_scale):
for ratio_index, ratio in enumerate(anchor_ratio):
list_h_on_image.append(scale / math.sqrt(ratio))
list_w_on_image.append(scale * math.sqrt(ratio))
global_index += 1
# shape info:
# y_on_image, x_on_image: layers_shapes[0] * layers_shapes[1]
# h_on_image, w_on_image: num_anchors_along_depth
return tf.expand_dims(y_on_image, axis=-1), tf.expand_dims(x_on_image, axis=-1), \
tf.constant(list_h_on_image, dtype=tf.float32), \
tf.constant(list_w_on_image, dtype=tf.float32), num_anchors_along_depth, num_anchors_along_spatial
-------------------------------------------------------------------
将入参代入后：
with tf.name_scope('get_layer_anchors'):
x_on_layer, y_on_layer = tf.meshgrid(tf.range(38), tf.range(38))
# x_on_layer = [
# [0,1,2,...,37],
# [0,1,2,...,37],
# ...
# [0,1,2,...,37],
# ]
# x_on_layer 共38行
#
# y_on_layer = [
# [0,0,0,...,0],
# [1,1,1,...,1],
# ...
# [37,37,37,...,37],
# ]
# x_on_layer 共38列
#
y_on_image = (tf.cast(y_on_layer, tf.float32) + offset) * layer_step / self._img_shape[0]
x_on_image = (tf.cast(x_on_layer, tf.float32) + offset) * layer_step / self._img_shape[1]
# tf.cast(y_on_layer, tf.float32) 该步骤是将坐标转为float类型，方便计算
# (tf.cast(y_on_layer, tf.float32) + offset) * layer_step 该步骤是计算映射回原图后的坐标
# / self._img_shape[0] 该步骤是将坐标归一化
num_anchors_along_depth = len(anchor_scale) * len(anchor_ratio) + len(extra_anchor_scale)
num_anchors_along_spatial = layer_shape[1] * layer_shape[0]
# num_anchors_along_depth = 4
# num_anchors_along_spatial = 38*38 = 1524
list_h_on_image = []
list_w_on_image = []
global_index = 0
# for square anchors
# extra_anchor_scale = (0.1414,)
for _, scale in enumerate(extra_anchor_scale):
list_h_on_image.append(scale)
list_w_on_image.append(scale)
global_index += 1
# list_h_on_image = [0.1414]
# list_w_on_image = [0.1414]
# global_index = 1
# for other aspect ratio anchors
# anchor_scale = (0.1,)
# anchor_ratio = (1., 2., .5)
for scale_index, scale in enumerate(anchor_scale):
for ratio_index, ratio in enumerate(anchor_ratio):
list_h_on_image.append(scale / math.sqrt(ratio)) # 0.1/sqrt(1)
list_w_on_image.append(scale * math.sqrt(ratio))
global_index += 1
# list_h_on_image = [0.1414, 0.1, 0.0707, x, x]
# list_w_on_image = [0.1414, y, y, y, y]
# global_index = 4
# shape info:
# y_on_image, x_on_image: layers_shapes[0] * layers_shapes[1]
# h_on_image, w_on_image: num_anchors_along_depth
return tf.expand_dims(y_on_image, axis=-1), tf.expand_dims(x_on_image, axis=-1), \
tf.constant(list_h_on_image, dtype=tf.float32), \
tf.constant(list_w_on_image, dtype=tf.float32), num_anchors_along_depth, num_anchors_along_spatial
# x_on_layer = [
# [0,1,2,...,37],
# [0,1,2,...,37],
# ...
# [0,1,2,...,37],
# ]
# x_on_layer 共38行
#
# y_on_layer = [
# [0,0,0,...,0],
# [1,1,1,...,1],
# ...
# [37,37,37,...,37],
# ]
x_on_layer.shape = (38,38)
y_on_layer.shape = (38,38)
tf.expand_dims(y_on_image, axis=-1).shape = (38,38,1)
tf.expand_dims(x_on_image, axis=-1).shape = (38,38,1)
list_h_on_image = [0.1414, 0.1, 0.0707, x, x]
list_w_on_image = [0.1414, y, y, y, y]
num_anchors_along_depth = 4
num_anchors_along_spatial = 38*38 = 1444
anchors_this_layer = tf.expand_dims(y_on_image, axis=-1),
tf.expand_dims(x_on_image, axis=-1),
list_h_on_image,
list_w_on_image,
num_anchors_along_depth,
num_anchors_along_spatial
-------------------------------------------------------------------
all_anchors.append(anchors_this_layer[:-2]) = [six elements]
elment = (expand_dim_y_on_image(38,38,1),
expand_dim_x_on_image(38,38,1),
list_h_on_image(4,),
list_h_on_image(4,))
all_num_anchors_depth.append(anchors_this_layer[-2]) = [six elements]
elment = 4
all_num_anchors_spatial.append(anchors_this_layer[-1]) = [six elements]
elment = feature map size.
第二段
【参数】：
all_anchors.append(anchors_this_layer[:-2]) = [six elements]
elment = (expand_dim_y_on_image(38,38,1),
expand_dim_x_on_image(38,38,1),
list_h_on_image(4,),
list_h_on_image(4,))
all_num_anchors_depth.append(anchors_this_layer[-2]) = [six elements]
elment = 4
all_num_anchors_spatial.append(anchors_this_layer[-1]) = [six elements]
elment = feature map size
【代码段】：
all_anchors, all_num_anchors_depth, all_num_anchors_spatial = anchor_creator.get_all_anchors()
num_anchors_per_layer = []
for ind in range(len(all_anchors)):
num_anchors_per_layer.append(all_num_anchors_depth[ind] * all_num_anchors_spatial[ind])
anchor_encoder_decoder = anchor_manipulator.AnchorEncoder(allowed_borders = [1.0] * 6,
positive_threshold = FLAGS.match_threshold,
ignore_threshold = FLAGS.neg_threshold,
prior_scaling=[0.1, 0.1, 0.2, 0.2])
【过程参数】：
num_anchors_per_layer = [38*38*4, ] .shape = 6
