def generate_anchors_pre(height,
width,
feat_stride,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2),
frame_scale=1.0):
""" A wrapper function to generate anchors given different scales
Also return the number of anchors in variable 'length'
"""
anchor_scales = np.array(anchor_scales)*frame_scale
anchor_ratios = np.array(anchor_ratios)
anchors = generate_anchors(
ratios=anchor_ratios, scales=anchor_scales)
A = anchors.shape[0]
shift_x = np.arange(0, width) * feat_stride
shift_y = np.arange(0, height) * feat_stride
#Meshgrid output as xy (N2, N1)
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)):
""" A wrapper function to generate anchors given different scales
Also return the number of anchors in variable 'length'
### Params:
* `height`: Height of the feature map
* `width`: Width of the feature map
* `feat_stride`: feature stride
* `anchor_scales`: default is (8, 16, 32) which means (8^2, 16^2, 32^2)
* `anchor_ratios`: default is (0.5, 1, 2)
### Returns:
* `anchors`: a float 2D array with shape of [K x A, 4], where K = w x h,
A = len(anchor_ratio) x len(anchor_scale)
* `length`: length of the all anchors, value = (w x h) x (3 x 3)
"""
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)
# ravel(): Return a flattened array.
# shift length of [x1, y1, x2, y2]
shifts = np.vstack((shift_x.ravel(), shift_y.ravel(),
shift_x.ravel(), shift_y.ravel())).transpose()
K = shifts.shape[0] # w * h
# 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]) # K*A = (w*h) * (3*3)
return anchors, length
def generate_anchors_pre(height, width, feat_stride=16,
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'
"""
shift_x = tf.range(width) * feat_stride
shift_y = tf.range(height) * feat_stride
shift_x, shift_y = tf.meshgrid(shift_x, shift_y)
sx = tf.reshape(shift_x, shape=(-1,))
sy = tf.reshape(shift_y, shape=(-1,))
anchors = generate_anchors(ratios=np.array(anchor_ratios), scales=np.array(anchor_scales))
A = anchors.shape[0]
K = tf.multiply(width, height)
length = K * A
shifts = tf.transpose(tf.stack([sx, sy, sx, sy]))
shifts = tf.transpose(tf.reshape(shifts, shape=[1, K, 4]), perm=(1, 0, 2))
anchor_constant = tf.constant(anchors.reshape((1, A, 4)), dtype=tf.int32)
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_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, )) # - tf.cast(feat_stride[0]/2, tf.int32)
sy = tf.reshape(shift_y,
shape=(-1, )) # - tf.cast(feat_stride[0]/2, tf.int32)
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_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
# tf.range创建数字序列，x偏移的距离
shift_y = tf.range(height) * feat_stride # height
shift_x, shift_y = tf.meshgrid(shift_x, shift_y)
# tf.meshgrid用于从数组a和b产生网格
# 用法: [A,B]=Meshgrid(a,b)，生成size(b)Xsize(a)大小的矩阵A和B。
# 它相当于a从一行重复增加到size(b)行，把b转置成一列再重复增加到size(a)列。
sx = tf.reshape(shift_x, shape=(-1, )) #转换成列向量
sy = tf.reshape(shift_y, shape=(-1, ))
shifts = tf.transpose(tf.stack([sx, sy, sx, sy]))
# 把这四个列向量合并tf.transpose转置
K = tf.multiply(width, height)
# 两个矩阵中对应元素相乘
shifts = tf.transpose(tf.reshape(shifts, shape=[1, K, 4]), perm=(1, 0, 2))
# shape=(?,4,1),?应该是9
anchors = generate_anchors(ratios=np.array(anchor_ratios),
scales=np.array(anchor_scales))
A = anchors.shape[0]
# A=9
anchor_constant = tf.constant(anchors.reshape((1, A, 4)), dtype=tf.int32)
# 创建一个常数张量
length = K * A
anchors_tf = tf.reshape(tf.add(anchor_constant, shifts), shape=(length, 4))
# anchor_constant(9*4)是以[0 0 15 15]为中心生成的9个anchor，shift是偏移规则(9,4,1)
# tf.cast转换类型
return tf.cast(anchors_tf, dtype=tf.float32), length
def generate_anchors_pre_tf(clp_filter,
height,
width,
feat_stride=16,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2)):
# old code
# 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))
#todo: wn modified
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])) # 竖着排。。。。不，是横着
K = tf.multiply(width, height) # 原代码
# 将点云图展开 需要横向展开
# clp_filter = tf.placeholder(tf.int32, shape=[None, None]) # 参数获取
# clp_filter_flat = tf.reshape(tf.transpose(clp_filter), (K, -1)) # shape: [23*32, 1]
clp_filter_flat = tf.reshape(clp_filter, (K, -1)) # 修改后：横着
clp_filter_index = tf.where(clp_filter_flat)[:, 0]
# 获取筛选后的坐标点
shifts_filter = tf.gather(shifts, clp_filter_index) # shpae: [none, 4] 横着
K = tf.size(clp_filter_index, out_type=tf.int32)
shifts = tf.transpose(tf.reshape(shifts_filter, shape=[1, K, 4]),
perm=(1, 0, 2))
# todo: temp change
# shifts = tf.transpose(tf.stack([sx, sy, sx, sy])) # 竖着排
# 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)) # 生成9个anchors
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=16,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2)):
shift_x = tf.range(width) * feat_stride # width 张量[0,16,32,...]
shift_y = tf.range(height) * feat_stride # height 同上
#tf.meshgrid给定 N 个一维坐标数组 *args，返回 N 维坐标数组的列表输出，用于计算 N 维网格上的表达式。
shift_x, shift_y = tf.meshgrid(shift_x, shift_y)
# shift_x [[0,16,32]...]
# [0,16,32]...]
# .
# .
# .
# [0,16,32]...]]
#shift_y [[0, 0, 0,...]
# [16,16,16,...]
# [32,32,32,...]
# .
# .
# .]
sx = tf.reshape(shift_x, shape=(-1, )) #转换为一维张量
sy = tf.reshape(shift_y, shape=(-1, ))
# tf.stack将秩为 R 的张量列表堆叠成一个秩为 (R+1) 的张量
# tf.transpose 交换矩阵维度
shifts = tf.transpose(tf.stack([sx, sy, sx, sy]))
#shifts*T [[0,16,32,...,0,16,32,...]
# [0,0,0,...,16,16,16...]
# [0,16,32,...,0,16,32,...]
# [0,0,0,...,16,16,16...]]
#
# shift [[0,0,0,0]
# [16,0,16,0]
# [32,0,32,0]
# ...
# [0,16,0,16]
# ...
# ]
K = tf.multiply(width, height)
shifts = tf.transpose(tf.reshape(shifts, shape=[1, K, 4]), perm=(1, 0, 2))
# shifts.shape [K,1,4]
# 生成anchor
anchors = generate_anchors(ratios=np.array(anchor_ratios),
scales=np.array(anchor_scales))
#A为anchor个数
## 9个
A = anchors.shape[0]
#申明tf图中的常量
anchor_constant = tf.constant(anchors.reshape((1, A, 4)), dtype=tf.int32)
#anchor+偏移量
#length为总anchor数量
length = K * A
## 原始的9个anchor加偏移量，得到最终anchors
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'
'''
# 锚框记录的两个坐标点是左下角右上角
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
# ----------------- #
# x = np.array([0, 1, 2])
# y = np.array([0, 1])
# np.meshgrid()函数说明
# X, Y = np.meshgrid(x, y)
# X = [[0 1 2]
# [0 1 2]]
# Y = [[0 0 0]
# [1 1 1]]
# ------------------ # 生成图像每个点坐标位置
shift_x, shift_y = np.meshgrid(shift_x, shift_y)
#******************************#
# np.hstack() np.vstack()实例说明
# a = np.array([[1, 2],[2, 3],[3, 4]])
# b = np.array([[2, 3],[3, 4],[4, 5]])
# np.hstack((a,b)) =
# array([[1, 2, 2, 3],
# [2, 3, 3, 4],
# [3, 4, 4, 5]])
# np.vstack((a,b)) =
# array([[1, 2],
# [2, 3],
# [3, 4],
# [2, 3],
# [3, 4],
# [4, 5]])
# np.ravel() 解释
# a = arange(12).reshape(3,4) = [[ 0 1 2 3]
# [ 4 5 6 7]
# [ 8 9 10 11]]
# a.ravel() = [ 0 1 2 3 4 5 6 7 8 9 10 11]
#******************************#
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=16,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2)):
""" Generate anchor boxes
:param height: number of vertical anchor boxes that fit into image dimensions
:param width: number of vertical anchor boxes that fit into image dimensions
:param feat_stride: the size of the feature stide
:param anchor_scales: scales of anchors
:param anchor_ratios: aspect ratios of anchor boxes
:return:
"""
# Generate all of the horizontal and vertical positions for anchors.
shift_x = tf.range(width) * feat_stride
shift_y = tf.range(height) * feat_stride
# Create a meshgrid of the positions of all anchors
shift_x, shift_y = tf.meshgrid(shift_x, shift_y)
# Flatten meshgrids into 1D arrays
sx = tf.reshape(shift_x, shape=(-1, ))
sy = tf.reshape(shift_y, shape=(-1, ))
# Stack flat meshgrids together twice to be used as increments for anchors shaped like
# [x_min, y_min, x_max, y_max]
shifts = tf.transpose(tf.stack([sx, sy, sx, sy]))
# Find total number of shifts for anchors
K = tf.multiply(width, height)
# Permutate dimensions
shifts = tf.transpose(tf.reshape(shifts, shape=[1, K, 4]), perm=(1, 0, 2))
# Generates all of the anchors of all aspect ratios. Generates (num_ratios x num_scales)
# matrix of vectors containing [x_min, y_min, x_max, y_max]
anchors = generate_anchors(ratios=np.array(anchor_ratios),
scales=np.array(anchor_scales))
# Get the number of anchors
A = anchors.shape[0]
# shape of object storing anchors. In default case dim = (1, 9, 4)
anchor_constant = tf.constant(anchors.reshape((1, A, 4)), dtype=tf.int32)
# Number of all anchors for all shifts
length = K * A
# Initialize positions of all anchors, by incrememnting the base anchor with all points on the
# meshgrid
anchors_tf = tf.reshape(tf.add(anchor_constant, shifts), shape=(length, 4))
# return matrix of all anchors as well as number of all anchors
return tf.cast(anchors_tf, dtype=tf.float32), length
def generate_anchors_fixate(height, width, h_start, w_start, anchor_stride=(16,), anchor_sizes=(128, 256, 512), 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(anchor_stride, anchor_ratios, anchor_sizes)
A = anchors.shape[0]
shift_x = (np.arange(0, width) + w_start) * anchor_stride
shift_y = (np.arange(0, height) + h_start) * anchor_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(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]
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(height, width,
aspect_ratio=[0.5, 1.0, 2.0],
aspect_scale=[8, 16, 32],
feat_stride=16):
anchors = generate_anchors()
anchor_nums = anchors.shape[0]
width = np.arange(width)*feat_stride
height = np.arange(height)*feat_stride
shift_x, shift_y = np.meshgrid(width, height)
shifts = np.vstack((shift_x.ravel(), shift_y.ravel(), shift_x.ravel(), shift_y.ravel())).transpose()
print(shifts)
anchors = anchors.reshape(1, anchor_nums, 4) + shifts.reshape(1, shifts.shape[0], 4).transpose([1, 0, 2])
anchors = anchors.reshape(shifts.shape[0]*anchor_nums, 4).astype(np.float32)
print(anchors.shape)
print(anchors)
length = 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)):
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,
aspect_ratio=[0.5, 1.0, 2.0],
aspect_scale=[8, 16, 32],
feat_stride=16):
anchors = generate_anchors()
anchor_nums = anchors.shape[0]
width = np.arange(width) * feat_stride
height = np.arange(height) * feat_stride
shift_x, shift_y = np.meshgrid(width, height)
shifts = np.vstack((shift_x.ravel(), shift_y.ravel(), shift_x.ravel(),
shift_y.ravel())).transpose()
print(shifts)
anchors = anchors.reshape(1, anchor_nums, 4) + shifts.reshape(
1, shifts.shape[0], 4).transpose([1, 0, 2])
anchors = anchors.reshape(shifts.shape[0] * anchor_nums,
4).astype(np.float32)
print(anchors.shape)
print(anchors)
length = 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'
"""
# 1. 生成9个不同面积不同长宽比的窗口anchors，此时anchor有H, W两个参数
anchors = generate_anchors(ratios=np.array(anchor_ratios), scales=np.array(anchor_scales))
A = anchors.shape[0]
# 2. 根据height, width得到所有中心点坐标shifts
# todo: 修改中心点坐标与数量
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]
# 3. anchors中添加坐标信息，此时anchor有H, W, X, Y四个参数
# 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,
anchor_stride=(16, ),
anchor_sizes=(128, 256, 512),
anchor_ratios=(0.5, 1, 2)):
shift_x = tf.range(width) * anchor_stride # width
shift_y = tf.range(height) * anchor_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(anchor_stride, anchor_ratios, anchor_sizes)
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'
"""
#转到generate_anchors。py,得到一系列anchor[x1,y1,x2,y2]左下和右上坐标
anchors = generate_anchors(ratios=np.array(anchor_ratios),
scales=np.array(anchor_scales))
A = anchors.shape[0] #anchor的数量9个
#16格一平移
shift_x = np.arange(0, width) * feat_stride # feat_stride=[16,]
shift_y = np.arange(0, height) * feat_stride
shift_x, shift_y = np.meshgrid(shift_x,
shift_y) #生成网格图纵向复制和横向复制均为（w×h）（w×h）
#shift=[shift_x, shift_y, shift_x, shift_y] shift_x.ravel[1× W*W] shift_y.ravel[1× H*H]
#shift=[w*h, 4]
shifts = np.vstack((shift_x.ravel(), shift_y.ravel(), shift_x.ravel(),
shift_y.ravel())).transpose()
K = shifts.shape[0] # =w*h
# width changes faster, so here it is H, W, C
# transpose做轴对换0轴和1轴对换
# ---------------A=9 K=W*H----------------------------------
#[1,9,4]+[k,1,4]=[k,A,4]
#一个anchor生成4K个坐标，共扫图k次
anchors = anchors.reshape((1, A, 4)) + shifts.reshape((1, K, 4)).transpose(
(1, 0, 2))
#[K*A,4]，生成9K个anchor
anchors = anchors.reshape((K * A, 4)).astype(np.float32, copy=False)
#length=K*A
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)):
# 一个包装函数，用于生成给定不同比例的锚点 并返回anchors数量
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
# 接受两个一维数组生成两个二维矩阵 维度均为 y.shape x x.shape
# shift_x 每一行均为shift_x,shift_y每一列均为shift_y
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]
# 这里是 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=16,
anchor_scales=(8, 16, 32),
anchor_ratios=(0.5, 1, 2)):
#生成候选框的思路是计算基础框的9个范围，然后通过加法作为偏移，得到9K个候选框
shift_x = tf.range(width) * feat_stride #x偏移量
shift_y = tf.range(height) * feat_stride #y偏移量
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])) #增加一维
K = tf.multiply(width, height) #像素点总数
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)) #9个候选框
A = anchors.shape[0] #A=9,共9个候选框
anchor_constant = tf.constant(anchors.reshape((1, A, 4)), dtype=tf.int32)
length = K * A #共9×K个候选框
anchors_tf = tf.reshape(tf.add(anchor_constant, shifts), shape=(length, 4))
return tf.cast(anchors_tf, dtype=tf.float32), length
im_size_min = min(img.shape)
im_size_max = max(img.shape)
im_scale = float(args.scale) / float(im_size_min)
# Prevent the biggest axis from being more than MAX_SIZE
if np.round(im_scale * im_size_max) > args.max_scale:
im_scale = float(args.max_scale) / float(im_size_max)
img = cv2.resize(img, None, None, fx=im_scale, fy=im_scale, interpolation=cv2.INTER_LINEAR)
print("Scaled image size")
print(img.shape)
width = img.shape[1]
height = img.shape[0]
base_anchors = generate_anchors(base_height=11,
num_anchors=10,
anchor_width=16,
h_ratio_step=0.7)
heights = [x[3] - x[1] for x in base_anchors]
img = draw_anchors(img, heights, 16, (width // 2, height // 2))
img = draw_anchors(img, heights, 16, (100, 150))
img = draw_anchors(img, heights, 16, (width - 300, 150))
img = draw_anchors(img, heights, 16, (100, height - 150))
img = draw_anchors(img, heights, 16, (width - 300, height - 150))
cv2.namedWindow('test')
cv2.imshow('test', img)
cv2.waitKey()
if np.round(im_scale * im_size_max) > args.max_scale:
im_scale = float(args.max_scale) / float(im_size_max)
img = cv2.resize(img,
None,
None,
fx=im_scale,
fy=im_scale,
interpolation=cv2.INTER_LINEAR)
print("Scaled image size")
print(img.shape)
width = img.shape[1]
height = img.shape[0]
base_anchors = generate_anchors(base_height=11,
num_anchors=10,
anchor_width=16,
h_ratio_step=0.7)
heights = [x[3] - x[1] for x in base_anchors]
img = draw_anchors(img, heights, 16, (width // 2, height // 2))
img = draw_anchors(img, heights, 16, (100, 150))
img = draw_anchors(img, heights, 16, (width - 300, 150))
img = draw_anchors(img, heights, 16, (100, height - 150))
img = draw_anchors(img, heights, 16, (width - 300, height - 150))
cv2.namedWindow('test')
cv2.imshow('test', img)
cv2.waitKey()
