Focal Loss for Dense Object Detection
PDF: https://arxiv.org/pdf/1708.02002.pdf
PyTorch代码: https://github.com/shanglianlm0525/PyTorch-Networks

创新点:
1 提出的Focal loss改进原始的CE loss来解决正负样本不平衡问题
2 设计了一个简单的one-stage 目标检测网络结构 RetinaNet

1 Focal Loss

CE loss:

Balanced CE loss:

Focal loss:

focal loss所加的指数式系数可对正负样本对loss的贡献自动调节。当某样本类别比较明确些，它对整体loss的贡献就比较少；而若某样本类别不易区分，则对整体loss的贡献就相对偏大。

** α-balanced Focal loss:**

α-balanced Focal loss 对不同类别更加平衡,实验效果更好.

下图为Focal loss

2 RetinaNet网络结构:

RetinaNet本质上是由resnet+FPN+两个FCN子网络组成，

PyTroch代码:

import torch
import torch.nn as nn
import torchvision

def Conv3x3ReLU(in_channels,out_channels):
    return nn.Sequential(
        nn.Conv2d(in_channels=in_channels,out_channels=out_channels,kernel_size=3,stride=1,padding=1),
        nn.ReLU6(inplace=True)
    )

def locLayer(in_channels,out_channels):
    return nn.Sequential(
            Conv3x3ReLU(in_channels=in_channels, out_channels=in_channels),
            Conv3x3ReLU(in_channels=in_channels, out_channels=in_channels),
            Conv3x3ReLU(in_channels=in_channels, out_channels=in_channels),
            Conv3x3ReLU(in_channels=in_channels, out_channels=in_channels),
            nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=1, padding=1),
        )

def confLayer(in_channels,out_channels):
    return nn.Sequential(
        Conv3x3ReLU(in_channels=in_channels, out_channels=in_channels),
        Conv3x3ReLU(in_channels=in_channels, out_channels=in_channels),
        Conv3x3ReLU(in_channels=in_channels, out_channels=in_channels),
        Conv3x3ReLU(in_channels=in_channels, out_channels=in_channels),
        nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=1, padding=1),
    )

class RetinaNet(nn.Module):
    def __init__(self, num_classes=80, num_anchores = 9):
        super(RetinaNet, self).__init__()
        self.num_classes = num_classes
        resnet = torchvision.models.resnet50()
        layers = list(resnet.children())

        self.layer1 = nn.Sequential(*layers[:5])
        self.layer2 = nn.Sequential(*layers[5])
        self.layer3 = nn.Sequential(*layers[6])
        self.layer4 = nn.Sequential(*layers[7])

        self.lateral5 = nn.Conv2d(in_channels=2048, out_channels=256, kernel_size=1)
        self.lateral4 = nn.Conv2d(in_channels=1024, out_channels=256, kernel_size=1)
        self.lateral3 = nn.Conv2d(in_channels=512, out_channels=256, kernel_size=1)

        self.upsample4 = nn.ConvTranspose2d(in_channels=256, out_channels=256, kernel_size=4, stride=2, padding=1)
        self.upsample3 = nn.ConvTranspose2d(in_channels=256, out_channels=256, kernel_size=4, stride=2, padding=1)

        self.downsample6 = nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=2, padding=1)
        self.downsample6_relu = nn.ReLU6(inplace=True)
        self.downsample5 = nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=2, padding=1)

        self.loc_layer3 = locLayer(in_channels=256,out_channels=4*num_anchores)
        self.conf_layer3 = confLayer(in_channels=256,out_channels=self.num_classes*num_anchores)

        self.loc_layer4 = locLayer(in_channels=256, out_channels=4*num_anchores)
        self.conf_layer4 = confLayer(in_channels=256, out_channels=self.num_classes*num_anchores)

        self.loc_layer5 = locLayer(in_channels=256, out_channels=4*num_anchores)
        self.conf_layer5 = confLayer(in_channels=256, out_channels=self.num_classes*num_anchores)

        self.loc_layer6 = locLayer(in_channels=256, out_channels=4*num_anchores)
        self.conf_layer6 = confLayer(in_channels=256, out_channels=self.num_classes*num_anchores)

        self.loc_layer7 = locLayer(in_channels=256, out_channels=4*num_anchores)
        self.conf_layer7 = confLayer(in_channels=256, out_channels=self.num_classes*num_anchores)

        self.init_params()

    def init_params(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def forward(self, x):
        x = self.layer1(x)
        c3 =x = self.layer2(x)
        c4 =x = self.layer3(x)
        c5 = x = self.layer4(x)

        p5 = self.lateral5(c5)
        p4 = self.upsample4(p5) + self.lateral4(c4)
        p3 = self.upsample3(p4) + self.lateral3(c3)

        p6 = self.downsample5(p5)
        p7 = self.downsample6_relu(self.downsample6(p6))

        loc3 = self.loc_layer3(p3)
        conf3 = self.conf_layer3(p3)

        loc4 = self.loc_layer4(p4)
        conf4 = self.conf_layer4(p4)

        loc5 = self.loc_layer5(p5)
        conf5 = self.conf_layer5(p5)

        loc6 = self.loc_layer6(p6)
        conf6 = self.conf_layer6(p6)

        loc7 = self.loc_layer7(p7)
        conf7 = self.conf_layer7(p7)

        locs = torch.cat([loc3.permute(0, 2, 3, 1).contiguous().view(loc3.size(0), -1),
                    loc4.permute(0, 2, 3, 1).contiguous().view(loc4.size(0), -1),
                    loc5.permute(0, 2, 3, 1).contiguous().view(loc5.size(0), -1),
                    loc6.permute(0, 2, 3, 1).contiguous().view(loc6.size(0), -1),
                    loc7.permute(0, 2, 3, 1).contiguous().view(loc7.size(0), -1)],dim=1)

        confs = torch.cat([conf3.permute(0, 2, 3, 1).contiguous().view(conf3.size(0), -1),
                           conf4.permute(0, 2, 3, 1).contiguous().view(conf4.size(0), -1),
                           conf5.permute(0, 2, 3, 1).contiguous().view(conf5.size(0), -1),
                           conf6.permute(0, 2, 3, 1).contiguous().view(conf6.size(0), -1),
                           conf7.permute(0, 2, 3, 1).contiguous().view(conf7.size(0), -1),], dim=1)

        out = (locs, confs)
        return out

if __name__ == '__main__':
    model = RetinaNet()
    print(model)

    input = torch.randn(1, 3, 800, 800)
    out = model(input)
    print(out[0].shape)
    print(out[1].shape)