Towards Bridging Seamntic Gap to improve semantic segmentation

Introduction:
1：bridge/alleviate semantic gap/inconsistency解决/缓解语义鸿沟/不连续
2：aggregate multi-level features/ implement multi-level feature aggregation 聚合多层特征
3：capture multi-scale/abundant/more context information 捕获多尺度/大量的上下文信息
4：enhance/fuse shallow/deep features增强/融合浅层/深层特征
5：enhance high-level/low-level features增强高层/低层特征
6：shallow and deep features are complementary information浅层和深层特征是互补信息
7：extract multi-scale robust features 提取多尺度健壮特征
8：restore the resolution information恢复分辨率信息
9：features from shallow layers encode more detailed information ,features from deep layers encode more semantic information to distinguish different classes.来自浅层的特征会编码更多细节信息，来自深层的特征编码更多语义信息去辨别不同的类别。
10：shallow feature are low-level for the edges,lines,and corners, while deep features are high-level for measuring object characteristics ,classification,and scene parsing. We refer this kind of difference of semantic granularity as semantic gap.
对于边，线和角来说，浅层特征是低层的，对于衡量物体特征，分类和场景理解来说，深层特征是高层特征。我们讲这种不同定义为语义鸿沟。
11：feature from shallow layers contain/have more local region and edge /coarse information，feature from deep layers contain/have more discriminative/representative information.
浅层特征具有更多的局部区域和边界/粗略的信息，深层特征包含更多的具有辨别性/代表性的信息。
12：Introducing shallow features to deep features will bring some background ‘noisy’ features,which influences the feature robustness and may cause semantic inconsistency.
将浅层特征引入深层特征会带来背景噪声，这会影响特征的健壮性，并且会导致语义不连续。
13：one solution can be found to promote robust feature aggregation is …
用来促进健壮的特征聚合方法是…
14: produce multi-resolution features产生多个分辨率特征
15：build skip connection from the encoder features to the corresponding features in the decoder stream。
在编码器特征和解码器特征之间构造跳连接
16: features from different stages of the network contain information of various scales.
网络的不同阶段特征包含不同尺度的信息。
17：fuse the features from neighboring stages in a bottom up manner.
以从下而上的方式融合周围的阶段。
18：the semantic gap (existing between the shallow layers and deep layers) limits the performance of feature fusion.
位于浅层和深层之间的语义鸿沟限制了融合的效果。
19:Enlarging the receptive fields and capturing more context information can help improve the representation ability of features.增大感受野和捕捉更多的上下文信息可以帮助提升特征的表征能力。
20：Different from xxx which is used …,the xxx is used to do.
与使用xxx不同的是，xxx用来做xxx。
21: make the learning procedure stable.
使训练过程更加牢固。
Methods:
22：considering both efficiency and effectiveness
同时考虑效率和有效性
23：we use a 1x1 convoltional layer to reduce the input channel number to a small value.
我们使用1x1卷积去将输入通道数降低到一个小数值。
24：To generate precise segmenattion results, we propose to construct a xxx module .
为了产生准确的分割结果，我们提出去构造一个xxx模块。
25：we first employ xxx to capture the local region information. The following xxx is able to capture global context information on a xxx manner.
我们首先使用xxx去捕捉局部区域信息，接下来的xxx能够以xxx方式去捕捉全局上下文信息。
26：Using xxx, xxx can capture xxx.
使用xxx,xxx可以捕捉xxx。
27：While casaded convolutional operations can enlarge receptive fields and capture more global context information,they also result in a loss of boundary information.
尽管连续的卷积可以扩大感受野和捕捉全局上下文信息，但是这种操作也会导致边界信息的损失。
28: However ,the features in shallow layers not only contains boundary information but also contain texture information inside objects or stuff, which may negatively effect the robustness of high-level features.
浅层特征不仅包含边界信息而且包含纹理信息，它会对掉高层特征的鲁棒性有不好的影响。
29：we propose to extract boundary information by paying attention to the region that are not sailent in high-level features.我们提出通过更多的关注那些在高层特征不重要的区域来提取边界信息。
Experiments：
30：ablation experiments are carried out on the xxx dataset to evaluate the contribution of each part of our models.在xxx数据集上进行消融实验来评估模型的每一个部分的贡献。
31:Using xxx to enhance xxx,we obtain xxxmIoU,which is xxx better than that of xxx and xxx less than using xxx. It can be found that the semantic gap between xxx and xxx can be alleviate by introducing xxx.
使用xxx去增强xxx，我们获得xxxmIoU，比使用xxx高了xxx，比使用xxx低了xxx，这验证了通过引入xxx，xxx和xxx之间的语义鸿沟可以有效的被消除。