Introduction

These approaches(VGG, ResNet) exhibit clear limitations when it comes to dense prediction in tasks like dense depth or normal estimation

DeepLab使用空洞卷积来解决这个问题
First, it needs to perform convolutions on a large number of detailed (high-resolution) feature maps that usually have high-dimensional features, which are computationally expensive.
Second, dilated convolutions introduce a coarse sub-sampling of features, which potentially leads to a loss of important details.

另一种方式是利用网络中间层的信息
High-level semantic features helps the category recognition of image regions, while low-levelvisual features help to generate sharp, detailed boundaries for high-resolution prediction.

主要的贡献有：
We propose a multi-path refinement network (Re-fineNet) which exploits features at multiple levelsof abstraction for high-resolution semantic segmentation.

Our cascaded RefineNets can be effectively trained end-to-end, which is crucial for best prediction performance.

We propose a new network component we call“chained residual pooling” which is able to capture background context from a large image region.

We achieve an IoU score of83.4on the PASCAL VOC 2012 dataset

Related Work

FCN
DeepLab-CRF
CRF-RNN
Seg-Net
U-Net

Although there are a few existing work, how to effectively exploit middle layer features remains an open question.

Proposed Method

here is onlyone input for RefineNet-4
the output of RefineNet-4 and the ResNet block-3 are fed to RefineNet-3 as 2-path inputs.

RefineNet

RCU
The first part of each Re-fineNet block consists of an adaptive convolution set that mainly fine-tunes the pretrained ResNet weights for our task.

Multi-resolutin Fusison
All path inputs are then fused intoa high-resolution feature map by the multi-resolution fusion block, depicted in Fig. 3(c).

Chained Redisual Pooling
the proposed chained residualpooling aims to capture background context from a largeimage region.

Output convolutions
To reflect this behavior in the last RefineNet-1 block, we placetwo additional RCUs before the final softmax predictionstep.