[toc]
CNN Geometric 中文介绍
论文1: CNN Geometric
Convolutional neural network architecture for geometric matching
卷积神经网络结构用于几何匹配
I. Rocco, R. Arandjelović and J. Sivic. Convolutional neural network architecture for geometric matching. CVPR 2017 [website][arXiv]
架构：
阶段1：仿射变换 estimates an affine transformation
阶段2：薄板样条转换 thin-plate spline (TPS) transformation
Started:
- demo.py demonstrates the results on the ProposalFlow dataset (Proposal Flow Dataset 的示范结果)
- train.py is the main training script (训练入口)
- eval_pf.py evaluates on the ProposalFlow dataset (用于评估dataset)
Trained models
Using Streetview-synth dataset + VGG
Using Pascal-synth dataset + VGG
Using Pascal-synth dataset + ResNet-101
Streetview: 是通过对来自东京时间机器数据集[4]的图像应用合成变换生成的，该数据集包含了东京的谷歌街景图像
Pascal: created from the training set of Pascal VOC 2011 [16]
论文2: DGC-NET
DGC-Net: Dense Geometric Correspondence Network
稠密几何对应网络
架构：
四个组成部分：
- 特征金字塔（feature pyramid creator）siamese VGG16 双重连接；类似Vgg16的网络架构，进行特征提取
- 关联层 （correlation layer）：5 convolutional blocks (Conv-BN-ReLU) to estimate a 2D dense correspondence field
- 扭曲层（warp layer）：
- matchability译码器 （matchability decoder ）：It contains four convolutional layers outputting a probability map (parametrized as a sigmoid
