zhiwei2coder的博客

Pre-learning隐马尔科夫模型 Y={y1,y2,...,yn}Y = \{y_1, y_2,...,y_n\} 是一组随机变量，X={x1,x2,...,xn}X = \{x_1, x_2,...,x_n\} 是其观测变量，我们假设Y具有马尔科夫性，则X，Y的联合概率为 P(x1,x2,...,xn,y1,y2,...,yn)=P(y1)P(x1|y1)∏i=2nP(yi|yi−1)

contribution： 1）novel recurrent model, end-to-end unifying consistency check and disparity estimation 2) introduce a soft attention mechanism, make the network automatically focus on the unreliable ...

目录简介摘要论文主要内容A: 匹配代价计算特征提取（ResNet18）特征Volume构建相似度计算B: 匹配代价聚合C：视差计算实验简介Deep Stereo Matching with Explicit Cost Aggregation Sub-Architecture AAAI2018 ，KITTI2015目前（20180320...

（未结束）NotionEnvironment Installation https://askubuntu.com/questions/806160/how-to-check-if-libpng-and-png-is-installed →→\to sudo apt-get install libpng++-devContributions1) Presenting a ...

This proposed a new method to solve the textureless and specular surface problem, which is using object knowledge.

Notes of this paper which talk about unsupervised learning for stereo matchWe treat correspond problem as a transformation from the left to right image. First, we start with from the assumption tha...

[Paper note](Efficient Deep Learning for Stereo Matching)这篇文章介绍一种更加快速但是只损失少量精确度的双目视差计算方法。其使用的主要框架与MC-CNN类似，尤其是Fast 框架，直接对两个由Siamese network得到feature vectors进行inner product计算。然后对结果的预测采用多类分类的模型，类别即不同大...

https://www.youtube.com/watch?v=kmqSoinWqgQ

这篇文章介绍了一种新的双目视差计算框架，Cascade Residual Learning。近年来，在MC-CNN的启发下，视差计算算法成为了一种可以用学习解决的问题。文章总结了近年来基于CNN的视差算法，主要可以分为一下三种： 1. Matching cost Learning：这种方法主要将CNN用于匹配代价计算，辅之以一些经典的后处理方法，以得到比较好的视差图。主要代表有：Han ...

论文思想笔记结构论文思想使用ResNet提取特征用dilated convolution增大感受野用Pyramid pooling module，提取各个scale的特征，使之包含global的context信息用feature volume shift构建volume用stacked hourglass 3D CNN（encoder-deco...