deep

刘帅成博士，旷世成都研究院，该院2018年成立， 新加坡国立大学毕业 目录： 1基于传统算法 2基于深度学习 3基于硬件 Image aligment methods图像对齐 应用： 图片拼接：不同视角拍的建筑拼成更大的一个画面，全景图 去噪，防抖 HDR不同曝光的同一场景，选取不同的曝光区域 拼接 多摄像头协同工作，一个是长焦，一个是广角。 视频防抖。 挑战： 1 重复特征：大厦窗口纹理很像，重复纹理 2 弱纹理，无纹理：白色的墙面，海面天空 3 大前景的干扰：有很多路人，视角偏一点，遮挡后面的景象就变化了 4 夜景，噪声的干扰 5 大视差：视角角度差别比较大时 工业界里不是关注什么场景里效果最好能有多好，而是关心实用中最坏情况匹配的多坏 传统算法 Homography 3*3的矩阵 限定：只能描述平面运动，或者相机绕光心的纯旋转运动 依赖特征点； Meshflow 属于Mesh warps算法的一个分支。 给定两帧：t,t-1, 图像划分成小格子，算小格字中的特征点 一个好的特征点的分布应该是均匀的全覆盖在整个图片画面，（而不是SIFT这种会附着在梯度大的地方），在检测不到特征点的地方用另一个方程赋予它特征点，密集地区筛选一些 并且也可以解决图片中深度变化的地方 深度学习方法 Deep Homography 2016 Deep Image Homography

Resources Lessons Lesson 1 Notes Tim Lee (Tim’s GitHub repo) Lesson 2: Case Study - A world class image classifier for dogs and cats (err…, anything) Apil Tamang Lesson 2 Notes Tim Lee Lesson 3 Notes Tim Lee Lesson 4 Notes Tim Lee Blog Sites by Author Anand Saha Apil Tamang Blogs Written by (or recommended by) fastai Fellows Resnet Decoding the ResNet architecture Anand Saha Yet Another ResNet Tutorial (or not) Apil Tamang An Overview of ResNet and its Variants Vincent Fung Stuctured Deep Learning Structured Deep Learning Kerem Turgutlu (Masters’ student at USF) NLP Fine-tuned Language Models

【推荐】2019 Java 开发者跳槽指南.pdf(吐血整理) >>> 代码如下： deepCopy: function(obj) { return _.isObject(obj) ? JSON.parse(JSON.stringify(obj)) : obj; }, 转载如下： 粘贴过来格式就乱了： 原文在此：http://blog.sibo.me/2014/08/20/javascript-deep-clone.html 结论 深度克隆是一个复杂的任务，并没有一个完美的解决方案，要根据情况谨慎对待。 如果是简单场景，可以使用 JSON.parse(JSON.stringify(obj)) 完成。 如果自己完全知道对象的数据结构，手动克隆将是最稳妥也是最高效的方式。 参考 http://stackoverflow.com/questions/7914968/cloning-whats-the-fastest-alternative-to-json-parsejson-stringifyx What is the difference between a deep copy and a shallow copy? Explanation of Deep and Shallow Copying Deep Copy vs JSON Stringify / JSON Parse What

Fastai Deep Learning Course: Version 2 Dates of Course Deep Learning Version 2 (Oct 2017 to Apr 2018) Part 1: Oct - Dec 2017 Part 2: Mar - May 2018 Forums Discourse: part1-v2 Discourse: part1-v2 beginner Discourse: part2-v2 Deep Learning Coursework (Version 2) Part 1 v2 (released Jan 2018) Part 2 v2 (released May 2018) Deep Learning Part 1 Lesson 1 wiki Image Recognition Lesson 2 wiki CNNs Lesson 3 wiki Overfitting Lesson 4 wiki Embeddings Lesson 5 wiki NLP Lesson 6 wiki RNNs Lesson 7 wiki CNN Architecture Deep Learning, Part 2 Lesson 8 Object Detection Lesson 9 Single Shot Multibox Detector

Fastai Deep Learning Course: Version 1 Dates of Course (Version 1) Deep Learning (Oct 2016 to Apr 2017) Part 1: Oct - Dec 2016 Part 2: Mar - May 2017 Deep Learing Coursework (Version 1) Part 1 v1 Part 2 v1 Other fastai v1: Launch Announcement fastai_old (on GitHub) (old version) 来源： CSDN 作者： DrogoZhang 链接： https://blog.csdn.net/weixin_40400177/article/details/103606334

Geometry Meets Deep Learning 2017 2016 2015 2017 A Point Set Generation Network for 3D Object Reconstruction from a Single Image paper-CVPR2017 Visual Attribute Transfer through Deep Image Analogy paper-SIGGRAPH2017 FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks paper-CVPR2017 Inverse Compositional Spatial Transformer Networks paper-CVPR2017 Convolutional neural network architecture for geometric matching paper-CVPR2017 2016 Deep3D: Fully Automatic 2D-to-3D Video Conversion with Deep Convolutional Neural Networks paper-2016 View Synthesis by Appearance Flow paper-ECCV2016

Model-Compression-Papers Papers for neural network compression and acceleration. Partly based on link . Survey Recent Advances in Efficient Computation of Deep Convolutional Neural Networks , [arxiv '18] A Survey of Model Compression and Acceleration for Deep Neural Networks [arXiv '17] Quantization The ZipML Framework for Training Models with End-to-End Low Precision: The Cans, the Cannots, and a Little Bit of Deep Learning [ICML’17] Compressing Deep Convolutional Networks using Vector Quantization [arXiv’14] Quantized Convolutional Neural Networks for Mobile Devices [CVPR '16] Fixed-Point

给一手链接 https://www.luogu.com.cn/problem/P3379 算是lca的模板吧 #include<cstdio> #include<cstring> #include<algorithm> #include<cmath> #include<queue> using namespace std; const int M=1e6+7; int read(){ int ans=0,f=1,c=getchar(); while(c<'0'||c>'9'){if(c=='-') f=-1; c=getchar();} while(c>='0'&&c<='9'){ans=ans*10+(c-'0'); c=getchar();} return ans*f; } int deep[M],f[M][27],vis[M]; int n,m,S,first[M],cnt; struct node{int to,next;}e[2*M]; void ins(int x,int y){e[++cnt]=(node){y,first[x]}; first[x]=cnt;} void dfs(int x){ vis[x]=1; for(int i=1;(1<<i)<=deep[x];i++) f[x][i]=f[f[x][i-1]][i-1]; for(int i=first[x

模板网络流最大流 题目链接： https://www.luogu.com.cn/problem/P1343 源点为1，终点为n #include <bits/stdc++.h> using namespace std; const int maxn=200000+5; int last[maxn],nxt[maxn*2],to[maxn*2],wi[maxn*2],cnt=-1; int deep[maxn]; int s,t,n,m,x; void add(int u,int v,int w) { to[++cnt]=v; wi[cnt]=w; nxt[cnt]=last[u]; last[u]=cnt; return ; } bool bfs() { queue < int > Q; memset(deep,0,sizeof(deep)); while(!Q.empty()) Q.pop();//预处理，清空队列，deep Q.push(s); deep[s]=1;//加入源头 do{ int u=Q.front(); Q.pop(); for(int i=last[u];i!=-1;i=nxt[i]) if(wi[i]>0 && deep[to[i]]==0)//边有残量 点未被占领 { deep[to[i]]=deep[u]+1; Q.push(to[i]); }

emmm 题目也已经很清楚了 题目链接： https://www.luogu.com.cn/problem/P3376 #include <bits/stdc++.h> using namespace std; const int maxn=200000+5; int last[maxn],nxt[maxn*2],to[maxn*2],wi[maxn*2],cnt=-1; int deep[maxn]; int s,t,n,m; void add(int u,int v,int w) { to[++cnt]=v; wi[cnt]=w; nxt[cnt]=last[u]; last[u]=cnt; return ; } bool bfs() { queue < int > Q; memset(deep,0,sizeof(deep)); while(!Q.empty()) Q.pop(); Q.push(s); deep[s]=1; do{ int u=Q.front(); Q.pop(); for(int i=last[u];i!=-1;i=nxt[i]) if(wi[i]>0 && deep[to[i]]==0) { deep[to[i]]=deep[u]+1; Q.push(to[i]); } }while(!Q.empty()); if(deep[t]==0) return 0