RefineNet笔记

RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation

https://arxiv.org/pdf/1611.06612.pdf

Thought

1、多利用residual 结构， 包括广义的，利于优化
2、与 FPN 同时间出现，当时流行特征融合
3、Chained residual pooling 相当于自融合，何时有效？
4、https://blog.csdn.net/qq_36165459/article/details/78345269
贯彻作者认为“所有特征都是有用的”这一个观点。
RCU 相当于是抽取低层特征；多分辨率融合模块，相当于是抽取中间层特征，除了refine-net4 以外，也起到融合多尺度的特征的作用，这样可以保留因为下采样导致的信息丢失的缺陷；最后的链式残差池化，相当于是抽取高层特征

institution

1、subsampling 损失空间信息
2、deconv 不能 recover the low-level visual features (increase computation and memory, computation)

solution

1、encoder-decoder结构，feature fusion前后加了处理，更好恢复空间信息
2、大量使用残差结构：Short-range residual 局部， long-range residual 在block间，
Short-range residual connections refer to local shot-cut connections in one RCU or the residual pooling component, while long-range residual connections refer to the connection between RefineNet modules and the ResNet blocks.

RefineNet
RCU: batch-normalization layers are removed, channel is set to 512 for RefineNet-4 and 256 for the remaining ones
Multi-resolution fusion: output channel 以inputs中最小为准, output scale 以inputs中最大为准, 单input不用处理
Chained residual pooling: aims to capture background context from a large image region, stride 1, 类似于自融合
ReLU is important for the effectiveness of subsequent pooling operations and it also makes the model less sensitive to changes in the learning rate
fusion block : only linear transformation are employed

Experiments