CNN光流计算2--FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks

FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
CVPR2017
Code: https://github.com/lmb-freiburg/flownet2

本文是对 FlowNet 的改进，改进主要有三点：
1） 在训练层面，数据库的训练的顺序很重要 the schedule of presenting data during training is very important
2）组合使用多个CNN网络， develop a stacked architecture that includes warping of the second image with intermediate optical flow
3）设计了一个专门的网络来针对小的运动

1. Dataset Schedules
   
   

The best results are consistently achieved when first training on Chairs and only then fine-tuning on Things3D

We conjecture that the simpler Chairs dataset helps the network learn the general concept of color matching without developing possibly confusing priors for 3D motion and realistic lighting too early
先在简单的 Chairs dataset 上学习广义的颜色匹配，得到一个好的权值初始化，然后再在Things3D 学习3D运动和真实光照变化

1. Stacking Networks
   

各种网络组合，各种尝试，找到最优的组合啊

速度对比：

精度时间对比：