自动驾驶论文: DFNet: Semantic Segmentation on Panoramic Images with Dynamic Loss Weights and RFB

环视车位检测和车道线分割 DFNet: Semantic Segmentation on Panoramic Images with Dynamic Loss Weights and Residual Fusion Block
PDF: https://arxiv.org/pdf/1806.07226.pdf
PyTorch: https://github.com/shanglianlm0525/PyTorch-Networks

1 概述:

DFNet主要划分为三块: 基本模块(basic module)、特征提取模块(features extraction module)、细化模块(refinement module).

• 1 选择Densenet作为基本模块(basic module);
• 2 特征提取模块(features extraction module)由PSPNet提出的金字塔池模块(pyramid pooling module)后接卷积层和一个双线性上采样层组成.
• 3 细化模块(refinement module)使用卷积层和池化层组成的残差融合块(residual fusion block, RFB) 减轻上采样带来的噪声干扰以及辨别处于类别边界上的点的归属.

2 网络结构

3 创新点

• a 根据每个batch中的样本动态计算样本权重,权重计算公式如下:
  
  其中, $w_{i}$wi​是类别$w$w的权重, $c$c 是类别数, $i~(0,c)$i (0,c), $α$α 和 $β$β 分别是$w_{i}$wi​的上下界阈值,避免权重差异过大, $N$N 是batch中的全部像素数, $n_{i}$ni​ 是类别$i$i的像素数,

when $n_{i}$ni​ = 0, it means that the class i does not appear in this batch, we set the weight to 1. Because we need to increase the effect of small pixel number class on loss, so the smaller the $n_{i}$ni​ , the larger the wi is. N and c are constant, wi is just changed by $n_{i}$ni​ . When the $n_{i}$ni​ is the average number, $w_{i}$wi​ is calculated to be 1/2, the multiplicative coefficient of 1/2 is also used to decrease the $w_{i}$wi​ of large pixel number of class.

• b RFB中使用的结构(由卷积层和池化层组成), 实验表明结构(f)性能最好
  

4 实验效果展示