READING NOTE: ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation
来源:互联网 发布:长尾关键词怎么优化 编辑:程序博客网 时间:2024/04/29 20:54
TITLE: ENet A Deep Neural Network Architecture for Real-Time Semantic Segmentation
AUTHER: Adam Paszke, Abhishek Chaurasia, Sangpil Kim, Eugenio Culurciello
ASSOCIATION: University of Warsaw, Purdue University
FROM: arXiv:1606.02147
CONTRIBUTIONS
- A novel deep neural network architecture named ENet (efficient neural network) is propsed, which is quite efficient.
- A serie of designing strategies is discussed.
Design Choices
Network Architecture
Readers could refer to the paper to have a look at the network architecture. The network is inspired by ResNet structure, while the authers re-design it based on the specific task of semantic segmentation and their intuitions. The intial block and basic building block (bottlenect module) is shown in the following figure. After the intial block, a comparetively large encoder is constructed using the bottleneck module. On the other hand, a smaller decoder follows the encoder.
Design Strategy
- Feature map resolution: Small feature map resolution has two drawbacks 1) loss of finer information of edges and 2) smaller size compared with original image. The advantage is that small feature map resolution means larger receptive field and more context for the filters. The first problem is solved by adding more feature maps or unsampling technique.
- Early downsampling: Early downsampling is very helpful for boosting the efficiency of the network while persisting the performance. The idea is that visual information is highly redundant and that initial network layers should not directly contribute to classification but act as good feature extractors.
- Decoder size: In most previous works, the encoder and decoder have the same size, for example totally symmetric. In this work, the auther uses a larger encoder and a smaller decoder. The responsibility of encoder is to operate on smaller resolution data and provide for information processing and filtering. Instead, the role of the the decoder, is to upsample the output of the encoder, only fine-tuning the details.
- Nonlinear operations In this paper some interesting observations are carried out. The auther invetigates the effect of nonlinear operations by training the network using PReLU. All layers in the main branch behave nearly exactly like regular ReLUs, while the weights of PReLU inside bottleneck modules are negative. It means that typical identity shortcut in ResNet does not work well because of the limited depth of the network.
- Information-preserving dimensionality changes: A method of performing pooling operation in parallel with a convolution of stride 2 and concatenating resulting feature maps is used to guarentee efficiency and performance, just as the intial block shows.
- Factorizing filters: Using factorizing technique can achive a kernel of larger size while using less computations. In addition, deeper network and more times of non-linear operation helps simulate richer functions.
- Dilated convolutions: Dilated convolutions is a good way of maintaining feature resolution while boosting efficiency.
- Regularization: Spatial Dropout is used to prevent overfitting.
ADVANTAGES
- The network processes fast.
DISADVANTAGES
- The performance is comparatively inferior.
- READING NOTE: ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation
- READING NOTE: PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
- READING NOTE: Learning Deconvolution Network for Semantic Segmentation
- Decoupled deep neural network for semi-supervised semantic segmentation 笔记
- Reading Note: ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression
- READING NOTE: Understanding Convolution for Semantic Segmentation
- [Paper note] PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
- 语义分割,ICNet for Real-Time Semantic Segmentation
- A Simple Deep and Effective Neural Network for Semantic Role Labelling 论文阅读
- READING NOTE: SuperCNN: A Superpixelwise Convolutional Neural Network for Salient Object Detection
- READING NOTE: Aggregated Residual Transformations for Deep Neural Networks
- 论文阅读-《BlitzNet: A Real-Time Deep Network for Scene Understanding》
- 目标检测分割--BlitzNet: A Real-Time Deep Network for Scene Understanding
- SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation
- 【ICNet】ICNet for Real-Time Semantic Segmentation on High-Resolution Images
- 实时语义分割--ICNet for Real-Time Semantic Segmentation on High-Resolution Images
- 实时语义分割--ICNet for Real-Time Semantic Segmentation on High-Resolution Images
- A Real Time Radiosity Architecture
- android的View的大小写字母问题
- 回环变位
- ReactJS组件生命周期(ES5为例)
- 两个栈来模拟队列
- 循环链表(两个单链表组合成一个循环链表)
- READING NOTE: ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation
- 【Java_项目篇<1>】--JAVA实现坦克大战游戏--赋予敌人行动和攻击(五)
- XP系统连接不了WIFI的解决方案
- thinkphp中模版主题使用方法
- 是的,我更喜欢努力的自己
- Day19、双向链表、树(二叉树、遍历、删除结点)
- Vivado HLS从HDL到模型和C的描述(1):从运行一个程序开始
- 关于连续天数的几个思路。
- Android中使用SDcard进行文件的读取