（原）人体姿态识别PyraNet

转载请注明出处：

https://www.cnblogs.com/darkknightzh/p/12424767.html

论文：

Learning Feature Pyramids for Human Pose Estimation

https://arxiv.org/abs/1708.01101

第三方pytorch代码：

https://github.com/Naman-ntc/Pytorch-Human-Pose-Estimation

1. 整体结构

将hourglass的残差模块改为金字塔残差模块（白框），用于学习输入图像不同尺度的特征。

hourglass见https://www.cnblogs.com/darkknightzh/p/11486185.html。参考代码中的Hourglass内部也使用了PRM模块，而不是原始的Hourglass。

该算法在stacked hourglass的基础上更容易理解。

2. 金字塔残差模块PRM

论文给出了4中PRM（金字塔残差模块）的结构，最终发现PRM-B的效果最好，如下图所示。其中虚线代表同等映射，白色虚框代表该处无上采样或下采样。

3. 下采样

由于pooling下采样速度太快，下采样倍数最低为2，因而论文未使用pool。而是使用了fractional max-pooling的下采样方式，第c层的下采样率（论文中M=1，C=4）：

${{s}_{c}}={{2}^{-M\frac{c}{C}}},c=0,\cdots ,C,M\ge 1$

4. 训练及测试

训练阶段和其他姿态估计算法相似，都是估计热图，然后计算真值热图和估计热图的均方误差，如下

$L=\frac{1}{2}\sum\limits_{n=1}^{N}{\sum\limits_{k=1}^{K}{{{\left\| {{\mathbf{S}}_{k}}-{{{\mathbf{\hat{S}}}}_{k}} \right\|}^{2}}}}$

其中N为样本数量，K为关键点的数量（也即热图数量）

测试阶段，使用最后一个hourglass热图最大的score的位置作为关键点。由于该算法为自顶向下的姿态估计算法，输入网络的图像仅有一个人，因而最大score的位置即为对应的关键点。

${{\mathbf{\hat{z}}}_{k}}=\underset{\mathbf{p}}{\mathop{\arg \max }}\,{{\mathbf{\hat{S}}}_{k}}(\mathbf{p}),k=1,L,K$

5. 代码

PyraNet定义如下：

class PyraNet(nn.Module):
    """docstring for PyraNet"""
    def __init__(self, nChannels=256, nStack=4, nModules=2, numReductions=4, baseWidth=6, cardinality=30, nJoints=16, inputRes=256):
        super(PyraNet, self).__init__()
        self.nChannels = nChannels
        self.nStack = nStack
        self.nModules = nModules
        self.numReductions = numReductions
        self.baseWidth = baseWidth
        self.cardinality = cardinality
        self.inputRes = inputRes
        self.nJoints = nJoints

        self.start = M.BnReluConv(3, 64, kernelSize = 7, stride = 2, padding = 3)   # BN+ReLU+conv

        # 先通过两分支（1*1 conv+3*3 conv，1*1 conv+不同尺度特征之和+3*3 conv，这两分支求和，并使用1*1 conv升维），并在输入输出通道相等时，直接返回，否则使用1*1 conv相加
        self.res1 = M.ResidualPyramid(64, 128, self.inputRes//2, self.baseWidth, self.cardinality, 0)
        self.mp = nn.MaxPool2d(2, 2)
        self.res2 = M.ResidualPyramid(128, 128, self.inputRes//4, self.baseWidth, self.cardinality,)  # 先通过两分支，并在输入输出通道相等时，直接返回，否则使用1*1 conv相加
        self.res3 = M.ResidualPyramid(128, self.nChannels, self.inputRes//4, self.baseWidth, self.cardinality)  # 先通过两分支，并在输入输出通道相等时，直接返回，否则使用1*1 conv相加

        _hourglass, _Residual, _lin1, _chantojoints, _lin2, _jointstochan = [],[],[],[],[],[]

        for _ in range(self.nStack):   # 堆叠个数
            _hourglass.append(PyraNetHourGlass(self.nChannels, self.numReductions, self.nModules, self.inputRes//4, self.baseWidth, self.cardinality))
            _ResidualModules = []
            for _ in range(self.nModules):
                _ResidualModules.append(M.Residual(self.nChannels, self.nChannels))     # 输入和输出相等，只有3*(BN+ReLU+conv)
            _ResidualModules = nn.Sequential(*_ResidualModules)
            _Residual.append(_ResidualModules)
            _lin1.append(M.BnReluConv(self.nChannels, self.nChannels))        # BN+ReLU+conv
            _chantojoints.append(nn.Conv2d(self.nChannels, self.nJoints,1))   # 1*1 conv，维度变换
            _lin2.append(nn.Conv2d(self.nChannels, self.nChannels,1))         # 1*1 conv，维度变换
            _jointstochan.append(nn.Conv2d(self.nJoints,self.nChannels,1))    # 1*1 conv，维度变换

        self.hourglass = nn.ModuleList(_hourglass)
        self.Residual = nn.ModuleList(_Residual)
        self.lin1 = nn.ModuleList(_lin1)
        self.chantojoints = nn.ModuleList(_chantojoints)
        self.lin2 = nn.ModuleList(_lin2)
        self.jointstochan = nn.ModuleList(_jointstochan)

    def forward(self, x):
        x = self.start(x)
        x = self.res1(x)
        x = self.mp(x)
        x = self.res2(x)
        x = self.res3(x)
        out = []

        for i in range(self.nStack):
            x1 = self.hourglass[i](x)
            x1 = self.Residual[i](x1)
            x1 = self.lin1[i](x1)
            out.append(self.chantojoints[i](x1))
            x1 = self.lin2[i](x1)
            x = x + x1 + self.jointstochan[i](out[i])     # 特征求和

        return (out)

ResidualPyramid定义如下：

class ResidualPyramid(nn.Module):
    """docstring for ResidualPyramid"""
    # 先通过两分支（1*1 conv+3*3 conv，1*1 conv+不同尺度特征之和+3*3 conv，这两分支求和，并使用1*1 conv升维），并在输入输出通道相等时，直接返回，否则使用1*1 conv相加
    def __init__(self, inChannels, outChannels, inputRes, baseWidth, cardinality, type = 1):
        super(ResidualPyramid, self).__init__()
        self.inChannels = inChannels
        self.outChannels = outChannels
        self.inputRes = inputRes
        self.baseWidth = baseWidth
        self.cardinality = cardinality
        self.type = type
        # PyraConvBlock：两分支，一个是1*1 conv+3*3 conv，一个是1*1 conv+不同尺度特征之和+3*3 conv，这两分支求和，并使用1*1 conv升维
        self.cb = PyraConvBlock(self.inChannels, self.outChannels, self.inputRes, self.baseWidth, self.cardinality, self.type)
        self.skip = SkipLayer(self.inChannels, self.outChannels)         # 输入和输出通道相等，则为None，否则为1*1 conv

    def forward(self, x):
        out = 0
        out = out + self.cb(x)
        out = out + self.skip(x)
        return out

PyraConvBlock如下：

class PyraConvBlock(nn.Module):
    """docstring for PyraConvBlock"""     # 两分支，一个是1*1 conv+3*3 conv，一个是1*1 conv+不同尺度特征之和+3*3 conv，这两分支求和，并使用1*1 conv升维
    def __init__(self, inChannels, outChannels, inputRes, baseWidth, cardinality, type = 1):
        super(PyraConvBlock, self).__init__()
        self.inChannels = inChannels
        self.outChannels = outChannels
        self.inputRes = inputRes
        self.baseWidth = baseWidth
        self.cardinality = cardinality
        self.outChannelsby2 = outChannels//2
        self.D = self.outChannels // self.baseWidth
        self.branch1 = nn.Sequential(   # 第一个分支，1*1 conv + 3*3 conv
                BnReluConv(self.inChannels, self.outChannelsby2, 1, 1, 0),           # BN+ReLU+conv
                BnReluConv(self.outChannelsby2, self.outChannelsby2, 3, 1, 1)        # BN+ReLU+conv
            )
        self.branch2 = nn.Sequential(   # 第二个分支，1*1 conv + 3*3 conv
                BnReluConv(self.inChannels, self.D, 1, 1, 0),                        # BN+ReLU+conv
                BnReluPyra(self.D, self.cardinality, self.inputRes),                 # BN+ReLU+不同尺度的特征之和
                BnReluConv(self.D, self.outChannelsby2, 1, 1, 0)                     # BN+ReLU+conv
            )
        self.afteradd = BnReluConv(self.outChannelsby2, self.outChannels, 1, 1, 0)   # BN+ReLU+conv

    def forward(self, x):
        x = self.branch2(x) + self.branch1(x)                                        # 两个分支特征之和
        x = self.afteradd(x)                                                         # 1*1 conv进行升维
        return x

BnReluPyra如下

class BnReluPyra(nn.Module):
    """docstring for BnReluPyra"""     # BN + ReLU + 不同尺度的特征之和
    def __init__(self, D, cardinality, inputRes):
        super(BnReluPyra, self).__init__()
        self.D = D
        self.cardinality = cardinality
        self.inputRes = inputRes
        self.bn = nn.BatchNorm2d(self.D)
        self.relu = nn.ReLU()
        self.pyra = Pyramid(self.D, self.cardinality, self.inputRes)     # 将不同尺度的特征求和

    def forward(self, x):
        x = self.bn(x)
        x = self.relu(x)
        x = self.pyra(x)
        return x

Pyramid如下：

class Pyramid(nn.Module):
    """docstring for Pyramid"""     # 将不同尺度的特征求和
    def __init__(self, D, cardinality, inputRes):
        super(Pyramid, self).__init__()
        self.D = D
        self.cardinality = cardinality     # 论文中公式3的C，金字塔层数
        self.inputRes = inputRes
        self.scale = 2**(-1/self.cardinality)   # 金字塔第1层的下采样率，后面层在此基础上+1
        _scales = []
        for card in range(self.cardinality):
            temp = nn.Sequential(    # 下采样 + 3*3 conv + 上采样
                    nn.FractionalMaxPool2d(2, output_ratio = self.scale**(card + 1)),  # 每一层在第1层基础上+1的下采样率
                    nn.Conv2d(self.D, self.D, 3, 1, 1),
                    nn.Upsample(size = self.inputRes)#, mode='bilinear')   # 上采样到输入分辨率
                )
            _scales.append(temp)
        self.scales = nn.ModuleList(_scales)

    def forward(self, x):
        #print(x.shape, self.inputRes)
        out = torch.zeros_like(x)         # 初始化和输入大小一样的0矩阵
        for card in range(self.cardinality):
            out += self.scales[card](x)    # 将所有尺度的特征求和
        return out

PyraNetHourGlass如下：

class PyraNetHourGlass(nn.Module):
    """docstring for PyraNetHourGlass"""
    def __init__(self, nChannels=256, numReductions=4, nModules=2, inputRes=256, baseWidth=6, cardinality=30, poolKernel=(2,2), poolStride=(2,2), upSampleKernel=2):
        super(PyraNetHourGlass, self).__init__()
        self.numReductions = numReductions
        self.nModules = nModules
        self.nChannels = nChannels
        self.poolKernel = poolKernel
        self.poolStride = poolStride
        self.upSampleKernel = upSampleKernel

        self.inputRes = inputRes
        self.baseWidth = baseWidth
        self.cardinality = cardinality

        """ For the skip connection, a residual module (or sequence of residuaql modules)  """
        # ResidualPyramid：先通过两分支，并在输入输出通道相等时，直接返回，否则使用1*1 conv相加
        # Residual：输入和输出相等，只有3*(BN+ReLU+conv)
        Residualskip = M.ResidualPyramid if numReductions > 1 else M.Residual
        Residualmain = M.ResidualPyramid if numReductions > 2 else M.Residual
        _skip = []
        for _ in range(self.nModules):  # 根据numReductions确定使用金字塔还是3*(BN+ReLU+conv)
            _skip.append(Residualskip(self.nChannels, self.nChannels, self.inputRes, self.baseWidth, self.cardinality))
        self.skip = nn.Sequential(*_skip)

        """ First pooling to go to smaller dimension then pass input through
        Residual Module or sequence of Modules then  and subsequent cases:
            either pass through Hourglass of numReductions-1 or pass through Residual Module or sequence of Modules """
        self.mp = nn.MaxPool2d(self.poolKernel, self.poolStride)

        _afterpool = []
        for _ in range(self.nModules):   # 根据numReductions确定使用金字塔还是3*(BN+ReLU+conv)
            _afterpool.append(Residualmain(self.nChannels, self.nChannels, self.inputRes//2, self.baseWidth, self.cardinality))
        self.afterpool = nn.Sequential(*_afterpool)

        if (numReductions > 1):     # 嵌套调用本身
            self.hg = PyraNetHourGlass(self.nChannels, self.numReductions-1, self.nModules, self.inputRes//2, self.baseWidth,
                                       self.cardinality, self.poolKernel, self.poolStride, self.upSampleKernel)
        else:
            _num1res = []
            for _ in range(self.nModules):    # 根据numReductions确定使用金字塔还是3*(BN+ReLU+conv)
                _num1res.append(Residualmain(self.nChannels,self.nChannels, self.inputRes//2, self.baseWidth, self.cardinality))
            self.num1res = nn.Sequential(*_num1res)  # doesnt seem that important ?

        """ Now another Residual Module or sequence of Residual Modules """
        _lowres = []
        for _ in range(self.nModules):    # 根据numReductions确定使用金字塔还是3*(BN+ReLU+conv)
            _lowres.append(Residualmain(self.nChannels,self.nChannels, self.inputRes//2, self.baseWidth, self.cardinality))
        self.lowres = nn.Sequential(*_lowres)

        """ Upsampling Layer (Can we change this??????) As per Newell's paper upsamping recommended  """
        self.up = nn.Upsample(scale_factor = self.upSampleKernel)     # 将高和宽扩充，实现上采样

    def forward(self, x):
        out1 = x
        out1 = self.skip(out1)             # 根据numReductions确定使用金字塔还是3*(BN+ReLU+conv)
        out2 = x
        out2 = self.mp(out2)               # 根据numReductions确定使用金字塔还是3*(BN+ReLU+conv)
        out2 = self.afterpool(out2)
        if self.numReductions>1:
            out2 = self.hg(out2)           # 嵌套调用本身
        else:
            out2 = self.num1res(out2)      # 根据numReductions确定使用金字塔还是3*(BN+ReLU+conv)
        out2 = self.lowres(out2)           # 根据numReductions确定使用金字塔还是3*(BN+ReLU+conv)
        out2 = self.up(out2)               # 升维

        return out2 + out1                 # 求和

Residual如下：

class Residual(nn.Module):
    """docstring for Residual"""     # 输入和输出相等，只有3*(BN+ReLU+conv)；否则输入通过1*1conv结果和3*(BN+ReLU+conv)求和
    def __init__(self, inChannels, outChannels, inputRes=None, baseWidth=None, cardinality=None, type=None):
        super(Residual, self).__init__()
        self.inChannels = inChannels
        self.outChannels = outChannels
        self.cb = ConvBlock(self.inChannels, self.outChannels)      # 3 * (BN+ReLU+conv) 其中第一组降维，第二组不变，第三组升维
        self.skip = SkipLayer(self.inChannels, self.outChannels)    # 输入和输出通道相等，则为None，否则为1*1 conv

    def forward(self, x):
        out = 0
        out = out + self.cb(x)
        out = out + self.skip(x)
        return out