Segmentation

pixel-wise labelling
干货 | 一文概览主要语义分割网络，FCN、UNet、SegNet、DeepLab 等等等等应有尽有

FCN (CVPR 2015)

Fully Convolutional Networks for Semantic Segmentation
final output low-resolution feature map(heatmap) as segmentation, upsampling with deconvolution backbone VGG16

SegNet (CVPR 2015)

SegNet: A Deep Convolutional Encoder-Decoder Architecture for Robust Semantic Pixel-Wise Labelling SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation Project unpooling: the decoder uses pooling indices computed in the max-pooling step of the corresponding encoder to perform non-linear upsampling

U-net (MICCAI 2015)

U-net

Mask R-CNN (ICCV 2017)

Mask R-CNN
extends Faster R-CNN by adding a branch for predicting segmentation masks on each Region of Interest


code: facebookresearch/Detectron

DeepLab

DeepLab v1 (ICLR 2015)

Semantic Image Segmentation with deep convolutional nets and fully connected CRFs

1. Dilated Convolution, aka Atrous Convolution
2. CRF post-process

DeepLab v2 (TPAMI 2017)

DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs
Atrous Spatial Pyramid Pooling, ASPP

DeepLab v3 (2017)

Rethinking Atrous Convolution for Semantic Image Segmentation

• MultiGrid
• Image-level feature

PSPNet (CVPR 2017)

Pyramid Scene Parsing Network
Some said that it is difficult to re-product its result.

RefineNet (CVPR 2017)

RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation
provides multiple paths over which information from different resolutions and via potentially long-range connections

Multi-Path Refinement

divide pre-trained ResNet into 4 blocks and employ a 4-cascaed architecture with 4 RefineNet units.

Residual Conv Unit + Multi-resolution Fusion

Chained residual pooling

ICNet (ECCV 2018)

ICNet for Real-Time Semantic Segmentation on High-Resolution Images
real-time
based on PSPNet

Cascade Feature Fusion, CFF

DANet

Dual attention network for scene segmentation (CVPR 2019)
pyTorch 1.4.0

integrate local features with their global dependencies based on the self-attention mechanism.

PAM

The position attention module selectively aggregates the feature at each position by a weighted sum of the features at all positions. Similar features would be related to each other regardless of their distances.

CAM

Meanwhile, the channel attention module selectively emphasizes interdependent channel maps by integrating associated features among all channel maps.