2020 最新 深度学习 图像分割技术 串烧

由于这两年深度学习图像分割技术出现大量有趣和前瞻的分割技术，现在这里做一篇回顾性的总结，以提升学习效率。涉及语义分割、实例分割等。

具体技术包罗：

1.卷积网络家族
2.编解码器家族
3.多尺度金字塔结构
4.R-CNN模型做实例分割
5.膨胀卷积和Deeplab系列
6.RNN
7.注意力机制
8.对抗网络GAN

神经网络分类

CNNs、

RNN (LSTM)

编解码器
Encoder-Decoder 和 Auto-Encoder模型

对抗网络

迁移学习

模型

FCN

Skip-connection

ParseNet

Encoder Decoder

SegNet


HRNet


U-Net


V-Net

Multi-layer perceptron (MLPs)


PSPN


R-CNN based models for Instance Segmentation

Mask R-CNN

MaskLab



Dilated Convolutional Models and Deeplab Family

DeepLab

DeepLab v3+

Attention-Based Models
Attention-based semantic segmentation model


The reverse attention network for segmentation


The dual attention netowrk for semantic segmentation


Generative Models and Adversarial Training


A semi-supervised segmentation framework


Adversarial network with Multi-scale L1 loss


EncNet
RefineNet
Seednet
Feedforward-Net
BoxSup
Graph-convolutional net
Wide ResNet
Exfuse
DIS dual image segmentation
FoveaNet
Ladder DenseNet
BiSeNet
SPGNet
Gated shape CNNs
AC-Net
DSSPN
SGR
SAC
UperNet

主流公开数据集

2D

1. PASCAL Visual Object Classed (VOC)
2. PASCAL context
3. Microsoft Common Objects in Context (MS COCO)
4. Cityscapes
5. ADE20K / MIT Scene Parsing (SceneParse150)
6. SiftFlow
7. Stanford background
8. Berkeley Segmentation Dataset (BSD)
9. Youtube-Ojects
10. KITTI

2.5D
11. NYU-D
12. SUN-3D
13. SUN RGB-D
14. UW RGB-D Object Dataset
15. ScanNet

3D
16. Stanford 2D-3D
17. ShapeNet Core
18. Sydney Urban Objects Dataset

主流评价指标

Pixel accuracy

Mean Pixel Accuracy (MPA)

IOU / Jaccard Index /Mean-IOU

Precision / Recall / F1

Dice

模型得分汇总