def generate_anchors_pre(height,
width,
feat_stride,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2)):
""" A wrapper function to generate anchors given different scales
Also return the number of anchors in variable 'length'
"""
"""生成anchor的预处理方法，generate_anchors方法就是直接产生各种大小的anchor box，generate_anchors_pre方法
是把每一个anchor box对应到原图上
feat_stride: 经过VGG或者ZF后特征图相对于原图的在长或者宽上的缩放倍数，也就是说height和width对应于特征图长宽
anchor_scales：anchor尺寸
anchor_ratios: anchor长宽比
"""
anchors = generate_anchors(ratios=np.array(anchor_ratios),
scales=np.array(anchor_scales))
A = anchors.shape[0] #A=9
# print("A:",A)
shift_x = np.arange(0, width) * feat_stride
shift_y = np.arange(0, height) * feat_stride
shift_x, shift_y = np.meshgrid(shift_x, shift_y)
shifts = np.vstack((shift_x.ravel(), shift_y.ravel(), shift_x.ravel(),
shift_y.ravel())).transpose()
K = shifts.shape[0]
# print("K:",K)
# width changes faster, so here it is H, W, C
anchors = anchors.reshape((1, A, 4)) + shifts.reshape((1, K, 4)).transpose(
(1, 0, 2))
anchors = anchors.reshape((K * A, 4)).astype(np.float32, copy=False)
length = np.int32(anchors.shape[0])
return anchors, length
def generate_anchors_pre(height,
width,
feat_stride,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2)):
""" A wrapper function to generate anchors given different scales
Also return the number of anchors in variable 'length'
"""
# 根据规模和比例，产生基础的9个anchor，大小为8,16,32，所以乘以缩放比例，就变成了128,256,512
anchors = generate_anchors(ratios=np.array(anchor_ratios),
scales=np.array(anchor_scales))
A = anchors.shape[0] # A = 9
# 生成X和Y偏移量
shift_x = np.arange(0, width) * feat_stride
shift_y = np.arange(0, height) * feat_stride
shift_x, shift_y = np.meshgrid(shift_x, shift_y)
shifts = np.vstack((shift_x.ravel(), shift_y.ravel(), shift_x.ravel(),
shift_y.ravel())).transpose()
K = shifts.shape[0] # width * height
# width changes faster, so here it is H, W, C
anchors = anchors.reshape((1, A, 4)) + shifts.reshape((1, K, 4)).transpose(
(1, 0, 2))
anchors = anchors.reshape((K * A, 4)).astype(np.float32, copy=False)
length = np.int32(anchors.shape[0])
return anchors, length
def generate_anchors_pre(height,
width,
feat_stride,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2)):
""" A wrapper function to generate anchors given different scales
Also return the number of anchors in variable 'length'
"""
anchors = generate_anchors(ratios=np.array(anchor_ratios),
scales=np.array(anchor_scales))
A = anchors.shape[0] # A = 9
shift_x = np.arange(0, width) * feat_stride
shift_y = np.arange(0, height) * feat_stride
shift_x, shift_y = np.meshgrid(shift_x, shift_y)
shifts = np.vstack((shift_x.ravel(), shift_y.ravel(), shift_x.ravel(),
shift_y.ravel())).transpose()
K = shifts.shape[0] # 1872
# width changes faster, so here it is H, W, C
anchors = anchors.reshape((1, A, 4)) + shifts.reshape((1, K, 4)).transpose(
(1, 0, 2))
anchors = anchors.reshape((K * A, 4)).astype(np.float32, copy=False)
length = np.int32(anchors.shape[0])
# print("anchors", anchors.shape, height*feat_stride)
anchors_dis = np.zeros((K * A, 3))
anchors_dis[:, 0] = ((anchors[:, 2] + anchors[:, 0]) / 2 -
width * feat_stride / 2) * 0.01234 # x对应距离
anchors_dis[:, 1] = (anchors[:, 3] +
anchors[:, 1]) * 0.0008963 # y对应距离 0.13349
anchors_dis[:, 2] = (375 - anchors[:, 3]) * 0.13349 # z 以米为单位
anchors_dis = anchors_dis.astype(np.float32, copy=False)
# print(anchors_dis)
return anchors, anchors_dis, length # anchors (16848, 4) anchors_dis (16848, 3)
def generate_anchors_pre(height,
width,
feat_stride,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1,
2)): #anchor_ratios = 2 w:h = 1:2
""" A wrapper function to generate anchors given different scales
Also return the number of anchors in variable 'length'
"""
anchors = generate_anchors(
ratios=np.array(anchor_ratios),
scales=np.array(anchor_scales)) #9个基本框，以后通过feet_stride调节,现在只生成一个基本框
A = anchors.shape[0]
shift_x = np.arange(0, width) * feat_stride #中心点的位置，，看看能不能做点文章
shift_y = np.arange(
0, height) * feat_stride #尝试传递上一个中心点坐标，推算下一个中心点坐标，从而减少anchor生成的数目
shift_x, shift_y = np.meshgrid(shift_x, shift_y)
shifts = np.vstack((shift_x.ravel(), shift_y.ravel(), shift_x.ravel(),
shift_y.ravel())).transpose()
K = shifts.shape[0]
# width changes faster, so here it is H, W, C
anchors = anchors.reshape((1, A, 4)) + shifts.reshape((1, K, 4)).transpose(
(1, 0, 2))
anchors = anchors.reshape((K * A, 4)).astype(np.float32, copy=False)
length = np.int32(anchors.shape[0])
return anchors, length
def generate_anchors_pre(height,
width,
feat_stride,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2)):
"""
Return anchors and its length.
A wrapper function to generate anchors given different scales
Also return the number of anchors in variable 'length'
将anchor从特征图上映射到原图上
Parameters
----------
height, width: int32
anchor的高度、宽度;
feat_stride: list
放大倍数,VGG16时,feat_stride=[16, ];
anchor_scales, anchor_ratios: float
anchor大小和宽高比的设定值
Returns
-------
anchors : ndarray
映射到原图上的anchor，shape=((number of anchors) x width x height, 4)
length : int32
anchors.shape[0] = (number of anchors) x width x height
"""
anchors = generate_anchors(ratios=np.array(anchor_ratios),
scales=np.array(anchor_scales))
A = anchors.shape[0]
# A = anchor的数量 = len(anchor_ritio) * len(anchor_scale) = 9
# 将特征图的宽、高以16倍延申至原图：经VGG16后得到的特征图大小缩小了16倍，故延申以还原
shift_x = np.arange(0, width) * feat_stride
# shift_x = [0, 1*16, ……, width*16], shape=[width, ]
shift_y = np.arange(0, height) * feat_stride
# shift_y = [0, 1*16, ……, height*16], shape=[height, ]
# 生成原图网格点
shift_x, shift_y = np.meshgrid(shift_x, shift_y)
# 将shift_x, shift_y生成形状相同的矩阵, 形为(height, width)
shifts = np.vstack((shift_x.ravel(), shift_y.ravel(), shift_x.ravel(),
shift_y.ravel())).transpose()
# 将shift_x, shift_y拉平后堆叠，然后转置
# shift_x(), shift_y.ravel(): (width*height, ); shifts: (4, width*height)^T = (width*height, 4)
K = shifts.shape[0]
# K = width x height
# width changes faster, so here it is H, W, C
anchors = anchors.reshape((1, A, 4)) + shifts.reshape((1, K, 4)).transpose(
(1, 0, 2)) # shape = [K, A, 4], K = H x W, A = 9
anchors = anchors.reshape((K * A, 4)).astype(np.float32, copy=False)
length = np.int32(anchors.shape[0])
# length = (number of anchors) x width x height
return anchors, length
def _anchor_component(self, size_level1, size_level2, size_level3):
anchors_level1, anchors_level2, anchors_level3 = generate_anchors(size_level1, size_level2, size_level3, self._feat_stride)
if cfg.NUM_ANCHORS_LEVEL1 != 0:
self._anchors_level1 = torch.from_numpy(anchors_level1)
if cfg.NUM_ANCHORS_LEVEL2 != 0:
self._anchors_level2 = torch.from_numpy(anchors_level2)
if cfg.NUM_ANCHORS_LEVEL3 != 0:
self._anchors_level3 = torch.from_numpy(anchors_level3)
def generate_anchors_pre_tf(height,
width,
feat_stride,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2)):
"""
A wrapper function to generate anchors given different scales and image
sizes in tensorflow.
Note, since `anchor_scales` and `anchor_ratios` is in practice always
the same, the generate 'base anchors' are static and does we only need to
implement the shifts part in tensorflow which is depending on the image
size.
Parameters:
-----------
height: tf.Tensor
The hight of the current image as a tensor.
width: tf.Tensor
The width of the current image as a tensor.
feat_stride: tf.Tensor or scalar
The stride used for the shifts.
anchor_scales: list
The scales to use for the anchors. This is a static parameter and can
currently not be runtime dependent.
anchor_ratios: list
The ratios to use for the anchors. This is a static parameter and can
currently not be runtime dependent.
Returns:
--------
anchors: tf.Tensor
2D tensor containing all anchors, it has the shape (n, 4).
length: tf.Tensor
A tensor containing the length 'n' of the anchors.
"""
anchors = generate_anchors(ratios=np.array(anchor_ratios),
scales=np.array(anchor_scales))
# Calculate all shifts
shift_x = tf.range(0, width * feat_stride, feat_stride)
shift_y = tf.range(0, height * feat_stride, feat_stride)
shift_x, shift_y = tf.meshgrid(shift_x, shift_y)
shift_x = tf.reshape(shift_x, [-1, 1])
shift_y = tf.reshape(shift_y, [-1, 1])
shifts = tf.concat((shift_x, shift_y, shift_x, shift_y), 1)
# Combine all base anchors with all shifts
anchors = anchors[tf.newaxis] + tf.transpose(shifts[tf.newaxis], (1, 0, 2))
anchors = tf.cast(tf.reshape(anchors, (-1, 4)), tf.float32)
length = tf.shape(anchors)[0]
return anchors, length
def generate_anchors_pre(height, width, feat_stride, anchor_scales=(8, 16, 32), anchor_ratios=(0.5, 1, 2)):
anchors = generate_anchors(ratios=np.array(anchor_ratios), scales=np.array(anchor_scales))
A = anchors.shape[0]
shift_x = np.arange(0, width) * feat_stride
shift_y = np.arange(0, height) * feat_stride
shift_x, shift_y = np.meshgrid(shift_x, shift_y)
shifts = np.vstack((shift_x.ravel(), shift_y.ravel(), shift_x.ravel(), shift_y.ravel())).transpose()
K = shifts.shape[0]
# width changes faster, so here it is H, W, C
anchors = anchors.reshape((1, A, 4)) + shifts.reshape((1, K, 4)).transpose((1, 0, 2))
anchors = anchors.reshape((K * A, 4)).astype(np.float32, copy=False)
length = np.int32(anchors.shape[0])
return anchors, length
def generate_anchors_pre_tf(height, width, feat_stride,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2)):
"""
A wrapper function to generate anchors given different scales and image
sizes in tensorflow.
Note, since `anchor_scales` and `anchor_ratios` is in practice always
the same, the generate 'base anchors' are static and does we only need to
implement the shifts part in tensorflow which is depending on the image
size.
Parameters:
-----------
height: tf.Tensor
The hight of the current image as a tensor.
width: tf.Tensor
The width of the current image as a tensor.
feat_stride: tf.Tensor or scalar
The stride used for the shifts.
anchor_scales: list
The scales to use for the anchors. This is a static parameter and can
currently not be runtime dependent.
anchor_ratios: list
The ratios to use for the anchors. This is a static parameter and can
currently not be runtime dependent.
Returns:
--------
anchors: tf.Tensor
2D tensor containing all anchors, it has the shape (n, 4).
length: tf.Tensor
A tensor containing the length 'n' of the anchors.
"""
anchors = generate_anchors(ratios=np.array(anchor_ratios),
scales=np.array(anchor_scales))
# Calculate all shifts
shift_x = tf.range(0, width * feat_stride, feat_stride)
shift_y = tf.range(0, height * feat_stride, feat_stride)
shift_x, shift_y = tf.meshgrid(shift_x, shift_y)
shift_x = tf.reshape(shift_x, [-1, 1])
shift_y = tf.reshape(shift_y, [-1, 1])
shifts = tf.concat((shift_x, shift_y, shift_x, shift_y), 1)
# Combine all base anchors with all shifts
anchors = anchors[tf.newaxis] + tf.transpose(shifts[tf.newaxis], (1, 0, 2))
anchors = tf.cast(tf.reshape(anchors, (-1, 4)), tf.float32)
length = tf.shape(anchors)[0]
return anchors, length
def generate_anchors_pre_tf(height, width, feat_stride=16, anchor_scales=(8, 16, 32), anchor_ratios=(0.5, 1, 2)): shift_x = tf.range(width) * feat_stride # width shift_y = tf.range(height) * feat_stride # height shift_x, shift_y = tf.meshgrid(shift_x, shift_y) sx = tf.reshape(shift_x, shape=(-1,)) sy = tf.reshape(shift_y, shape=(-1,)) shifts = tf.transpose(tf.stack([sx, sy, sx, sy])) K = tf.multiply(width, height) shifts = tf.transpose(tf.reshape(shifts, shape=[1, K, 4]), perm=(1, 0, 2)) anchors = generate_anchors(ratios=np.array(anchor_ratios), scales=np.array(anchor_scales)) A = anchors.shape[0] anchor_constant = tf.constant(anchors.reshape((1, A, 4)), dtype=tf.int32) length = K * A anchors_tf = tf.reshape(tf.add(anchor_constant, shifts), shape=(length, 4)) return tf.cast(anchors_tf, dtype=tf.float32), length
def generate_anchors_pre(height, width, feat_stride, anchor_scales=(8, 16, 32), anchor_ratios=(0.5, 1, 2)):
""" A wrapper function to generate anchors given different scales
Also return the number of anchors in variable 'length'
"""
anchors = generate_anchors(ratios=np.array(anchor_ratios), scales=np.array(anchor_scales))
A = anchors.shape[0] #9
shift_x = np.arange(0, width) * feat_stride #feature map上的点对应回原输入图片的坐标
shift_y = np.arange(0, height) * feat_stride
shift_x, shift_y = np.meshgrid(shift_x, shift_y)
# np.reval()和np.flatten()两者的功能是一致的，将多维数组降为一维
# 但两者的区别是no.flatten()返回的是一份拷贝，对拷贝做的修改不影响原始数组，而np.reval()返回的是视图，修改时会影响原数组
shifts = np.vstack((shift_x.ravel(), shift_y.ravel(), shift_x.ravel(), shift_y.ravel())).transpose()
K = shifts.shape[0]
# width changes faster, so here it is H, W, C
#anchors坐标加上加上各个中心点的坐标，即得到各个中心点产生的所有anchors的坐标
anchors = anchors.reshape((1, A, 4)) + shifts.reshape((1, K, 4)).transpose((1, 0, 2))
anchors = anchors.reshape((K * A, 4)).astype(np.float32, copy=False) #一个点产生的9个anchor连在一起
length = np.int32(anchors.shape[0])
return anchors, length
def generate_anchors_pre_tf(height,
width,
feat_stride=16,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2)):
"""
根据feature map的尺寸来计算该feature map所有像素点（像素点的选取顺序是从第一行开始自左到右的方向，处理完一行后自上到下进入下一行）对应
到原始图片上的所有anchors。每个对应的anchors的个数由anchor_scales和anchor_ratios决定，比如anchor_scales有8,16,32三种扩大倍数，
anchor_ratios有0.5,1,2三种纵横比，则对应feature map每个点，映射回原始图片的感受野后的中心点，会生成3 × 3中组合，比如0.5纵横比会组合
8,16,32三种扩大倍数，所以对于每个点有9个anchors。
Args:
height: feature map的高
width: feature map的宽
feat_stride: 从原始图片到该feature map总共缩小的倍数，由每层的stride决定
anchor_scales: anchors的扩大倍数
anchor_ratios: anchors的纵横比
Returns:
feature map所有像素点对应到原始图中的anchors
"""
"""
首先获取feature map所有点对应到原始图片的点的坐标，x原始 = x(feature map) * feat_stride, y原始 = x(feature map) * feat_stride.
比如3 × 3的feature map对应到原始图片中所有的坐标点shifts如下：
Out[133]:
array([[[ 0, 0, 0, 0]],
[[16, 0, 16, 0]],
[[32, 0, 32, 0]],
[[ 0, 16, 0, 16]],
[[16, 16, 16, 16]],
[[32, 16, 32, 16]],
[[ 0, 32, 0, 32]],
[[16, 32, 16, 32]],
[[32, 32, 32, 32]]], dtype=int32)
能看到第一列和第三列一致，第二列和第四列一致。一二列就组成了所有原始图中的点，增加三四列是为了方便计算anchor，因为一个anchor用左上角坐标
和右下角坐标来表示。
"""
shift_x = tf.range(width) * feat_stride # width
shift_y = tf.range(height) * feat_stride # height
shift_x, shift_y = tf.meshgrid(shift_x, shift_y)
sx = tf.reshape(shift_x, shape=(-1, ))
sy = tf.reshape(shift_y, shape=(-1, ))
shifts = tf.transpose(tf.stack([sx, sy, sx, sy]))
K = tf.multiply(width, height)
shifts = tf.transpose(tf.reshape(shifts, shape=[1, K, 4]), perm=(1, 0, 2))
"""
依据一个16 × 16的base anchor生成9个anchors（由scales和ratios组成），anchor_constant如下：
array([[ -84., -40., 99., 55.], 0.5纵横比，8倍扩大
[-176., -88., 191., 103.], 0.5纵横比，16倍扩大
[-360., -184., 375., 199.], 0.5纵横比，32倍扩大
[ -56., -56., 71., 71.], 1.0纵横比，8倍扩大
[-120., -120., 135., 135.], 1.0纵横比，16倍扩大
[-248., -248., 263., 263.], 1.0纵横比，32倍扩大
[ -36., -80., 51., 95.], 2.0纵横比，8倍扩大
[ -80., -168., 95., 183.], 2.0纵横比，16倍扩大
[-168., -344., 183., 359.]]) 2.0纵横比，32倍扩大
"""
anchors = generate_anchors(ratios=np.array(anchor_ratios),
scales=np.array(anchor_scales))
A = anchors.shape[0]
anchor_constant = tf.constant(anchors.reshape((1, A, 4)), dtype=tf.int32)
"""
让shifts与anchor_constant相加组成feature map对应到原始图片中的anchors。
在这里发现上边生成的shifts实际并非anchors的中心点。而是中心点偏左上一些的点。
总共anchors个数为: 像素点数K × 每个点的anchors数A
"""
length = K * A
anchors_tf = tf.reshape(tf.add(anchor_constant, shifts), shape=(length, 4))
return tf.cast(anchors_tf, dtype=tf.float32), length