补充文档的目录

WAMAWAMA · Jan 14, 2019 · 2ea52eb · 2ea52eb
1 parent eac8230
commit 2ea52eb
Show file tree

Hide file tree

Showing 58 changed files with 1,093 additions and 8 deletions.
diff --git a/图像分割/Automatic Portrait Segmentation for Image Stylization翻译(2016).md b/图像分割/Automatic Portrait Segmentation for Image Stylization翻译(2016).md
@@ -4,6 +4,34 @@
 > 论文: <http://xiaoyongshen.me/webpage_portrait/papers/portrait_eg16.pdf>
 > 翻译: <https://liqiang311.github.io/%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0/Automatic-Portrait-Segmentation-for-Image-Stylization-%E7%BF%BB%E8%AF%91%E5%AD%A6%E4%B9%A0/>
 
+* [Automatic Portrait Segmentation for Image Stylization](#automatic-portrait-segmentation-for-image-stylization)
+  * [Abstract](#abstract)
+  * [引言](#引言)
+  * [Related Work](#related-work)
+    * [Interactive Image Selection](#interactive-image-selection)
+    * [CNNs for Image segmentation](#cnns-for-image-segmentation)
+    * [Image Matting](#image-matting)
+    * [Semantic Stylization](#semantic-stylization)
+  * [Our Motivation and Approach](#our-motivation-and-approach)
+    * [Fully Convolutional Neutral Networks](#fully-convolutional-neutral-networks)
+    * [Understandings for Our Task](#understandings-for-our-task)
+    * [Our Approach](#our-approach)
+      * [Position Channels](#position-channels)
+      * [Shape Channel](#shape-channel)
+  * [Data and Model Training](#data-and-model-training)
+    * [Data Preparation](#data-preparation)
+    * [Model Training](#model-training)
+    * [Running Time for Training and Testing](#running-time-for-training-and-testing)
+  * [Results and Applications](#results-and-applications)
+    * [Quantitative and Visual Analysis](#quantitative-and-visual-analysis)
+    * [Post-processing](#post-processing)
+    * [User Study of Our Method](#user-study-of-our-method)
+    * [Automatic Segmentation for Image Stylization](#automatic-segmentation-for-image-stylization)
+    * [Other Applications](#other-applications)
+  * [Conclusions and Future Work](#conclusions-and-future-work)
+  * [补充资料](#补充资料)
+    * [homography](#homography)
+
 ## Abstract
 
 肖像是摄影和绘画的主要艺术形式。在大多数情况下，艺术家试图使主体从周围突出，例如，使其更亮或更锐利。在数字世界中，通过使用适合于图像语义的照相或绘画滤镜处理肖像图像，可以实现类似的效果。虽然存在用于描述主题的成功的用户引导方法，但缺乏全自动技术并且产生不令人满意的结果。我们的论文首先通过引入专用于肖像的新自动分割算法来解决这个问题。然后我们在此结果的基础上描述了几个利用我们的自动分割算法生成高质量肖像的肖像滤镜。

diff --git a/图像分割/DAG-Recurrent Neural Networks For Scene Labeling总结(2015).md b/图像分割/DAG-Recurrent Neural Networks For Scene Labeling总结(2015).md
@@ -4,6 +4,25 @@
 >
 > 这里应该指的是在CNN提取到的特征图上进行使用RNN.
 
+* [DAG-Recurrent Neural Networks For Scene Labeling(2015)](#dag-recurrent-neural-networks-for-scene-labeling2015)
+  * [关键点](#关键点)
+  * [简介](#简介)
+  * [相关工作](#相关工作)
+  * [网络结构](#网络结构)
+    * [DAG-RNN计算](#dag-rnn计算)
+    * [整体DAG-RNN](#整体dag-rnn)
+    * [损失函数](#损失函数)
+    * [处理类别不平衡](#处理类别不平衡)
+  * [测试](#测试)
+    * [基线模型](#基线模型)
+    * [实验使用的基本网络](#实验使用的基本网络)
+    * [SiftFlow Dataset](#siftflow-dataset)
+    * [CamVid Dataset](#camvid-dataset)
+    * [Barcelona Dataset](#barcelona-dataset)
+    * [Effects of DAG-RNNs to Per-class Accuracy](#effects-of-dag-rnns-to-per-class-accuracy)
+    * [Discussion of Modeled Dependency](#discussion-of-modeled-dependency)
+  * [总结](#总结)
+
 ## 关键点
 
 * 引入RNN来利用长期上下文依赖

diff --git a/图像分割/Deep Edge-Aware Saliency Detection论文翻译(2017).md b/图像分割/Deep Edge-Aware Saliency Detection论文翻译(2017).md
@@ -2,6 +2,32 @@
 
 > 论文: <https://arxiv.org/abs/1708.04366>
 
+* [Deep Edge-Aware Saliency Detection](#deep-edge-aware-saliency-detection)
+  * [摘要](#摘要)
+  * [前言](#前言)
+    * [Low-resolution output maps](#low-resolution-output-maps)
+    * [Missing handcrafted yet pivotal features:](#missing-handcrafted-yet-pivotal-features)
+    * [Archaic handling of multi-scale saliency:](#archaic-handling-of-multi-scale-saliency)
+    * [提出的方法](#提出的方法)
+  * [Joint Salient Edge and Saliency Detection](#joint-salient-edge-and-saliency-detection)
+    * [Reformulating Saliency Detection](#reformulating-saliency-detection)
+    * [Balanced Loss Function](#balanced-loss-function)
+    * [FCN for Edge-Aware Saliency](#fcn-for-edge-aware-saliency)
+    * [Effects of Reformulating Saliency Detection](#effects-of-reformulating-saliency-detection)
+  * [Integrating Deep and Handcrafted Features](#integrating-deep-and-handcrafted-features)
+    * [Handcrafted Saliency Features](#handcrafted-saliency-features)
+    * [Deep Multi-scale Saliency Feature Extraction](#deep-multi-scale-saliency-feature-extraction)
+    * [Deep-Shallow Model](#deep-shallow-model)
+    * [Context Module for Saliency Refinement](#context-module-for-saliency-refinement)
+  * [Experimental Results](#experimental-results)
+    * [Evaluation metrics](#evaluation-metrics)
+    * [消融实验](#消融实验)
+      * [数据集的不同](#数据集的不同)
+      * [Salient objects with diverse scales](#salient-objects-with-diverse-scales)
+      * [Different numbers of salient objects](#different-numbers-of-salient-objects)
+    * [执行时间](#执行时间)
+  * [Conclusions](#conclusions)
+
 ## 摘要
 
 由于深度学习架构，视觉显着性已经取得了长足的进步，然而，仍存在三个主要挑战，这些挑战阻碍了具有复杂构图，多个显着对象和不同尺度的显着对象的场景的检测性能。特别是，现有方法的输出图仍然保持较低的空间分辨率，由于步幅和汇集操作导致边缘模糊，网络经常忽略描述性统计和有可能补充显着性检测结果手工制作的先验，并且不同层次的深层特征主要保持不变, 等待被有效融合来处理多尺度显着对象。

diff --git a/图像分割/Deep Embedding Features for Salient Object Detection总结(2018).md b/图像分割/Deep Embedding Features for Salient Object Detection总结(2018).md
@@ -31,6 +31,25 @@
 
 这里的第6层是哪里来的?是前面的5C通道的特征输出汇总来的?
 
+* [Learning to Promote Saliency Detectors](#learning-to-promote-saliency-detectors)
+  * [Abstract](#abstract)
+  * [Introduction](#introduction)
+  * [Related works](#related-works)
+    * [BU](#bu)
+    * [TD](#td)
+    * [TD+BU](#tdbu)
+  * [The Proposed Method](#the-proposed-method)
+    * [The anchor network](#the-anchor-network)
+    * [Iterative testing scheme](#iterative-testing-scheme)
+    * [Pixel embedding](#pixel-embedding)
+    * [Region embedding](#region-embedding)
+  * [Experiments](#experiments)
+    * [评价标准](#评价标准)
+    * [Implementation details](#implementation-details)
+    * [Performance](#performance)
+  * [总结](#总结)
+  * [参考链接](#参考链接)
+
 ## Abstract
 
 The categories and appearance of salient objects varyfrom image to image, therefore, saliency detection is animage-specific task.  Due to lack of large-scale saliency training data, using deep neural networks (DNNs) with pre-training is difficult to precisely capture the image-specific saliency cues. To solve this issue, we formulate a **zero-shot learning** problem to promote existing saliency detectors.

diff --git a/图像分割/Deep Propagation Based Image Matting翻译(2018).md b/图像分割/Deep Propagation Based Image Matting翻译(2018).md
@@ -2,6 +2,28 @@
 
 > 论文: <https://www.ijcai.org/proceedings/2018/0139.pdf>
 
+* [Deep Propagation Based Image Matting](#deep-propagation-based-image-matting)
+  * [概要](#概要)
+  * [引言](#引言)
+  * [Related Work](#related-work)
+  * [Method](#method)
+    * [Deep Feature Extraction Module](#deep-feature-extraction-module)
+    * [Affinity Learning Module](#affinity-learning-module)
+    * [Matte Propagation Module](#matte-propagation-module)
+    * [Losses](#losses)
+    * [Implementation Detail](#implementation-detail)
+  * [Experimental Results](#experimental-results)
+    * [Dataset](#dataset)
+    * [Evaluation](#evaluation)
+  * [Conclusion](#conclusion)
+  * [重要参考](#重要参考)
+    * [论文](#论文)
+    * [拉普拉斯矩阵](#拉普拉斯矩阵)
+      * [介绍](#介绍)
+      * [定义](#定义)
+      * [性质](#性质)
+      * [例子](#例子)
+
 ## 概要
 
 在本文中，我们通过将深度学习引入学习alpha matte传播原理, 来提出一种基于深度传播的图像matting框架。我们的深度学习架构是深度特征提取模块，亲和力学习模块(an affinity learning module)和matte传播模块的串联。这三个模块都是不同的，可以通过端到端的训练流程进行共同优化。我们的框架通过学习适合于matte传播的深度图像表示，对于传播而言, 生成在像素的语义级别的成对相似性。它结合了深度学习和matte传播的强大功能，因此可以在准确性和训练复杂性方面超越先前最先进的matting技术，这可以通过我们基于两个基准matting数据集创建的243K图像上的实验结果得到验证.

diff --git a/...eep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs翻译(2016).md b/...eep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs翻译(2016).md
@@ -24,6 +24,28 @@
 
 ---
 
+* [DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs(2016)](#deeplab-semantic-image-segmentation-with-deep-convolutional-nets-atrous-convolution-and-fully-connected-crfs2016)
+  * [摘要](#摘要)
+  * [介绍](#介绍)
+    * [特征分辨率的下降 => atrous convolution](#特征分辨率的下降--atrous-convolution)
+    * [多尺度目标共存问题 => Atrous Spatial Pyramid Pooling (ASPP)](#多尺度目标共存问题--atrous-spatial-pyramid-pooling-aspp)
+    * [DCNN的不变形导致定位精度降低 => Conditional Random Field (CRF)](#dcnn的不变形导致定位精度降低--conditional-random-field-crf)
+  * [模型结构](#模型结构)
+  * [相关工作](#相关工作)
+    * [hand-crafted features](#hand-crafted-features)
+    * [DCNN提取特征](#dcnn提取特征)
+      * [cascade of bottom-up image segmentation](#cascade-of-bottom-up-image-segmentation)
+      * [DCNN features + segmentations](#dcnn-features--segmentations)
+      * [directly provide dense category-level pixel label](#directly-provide-dense-category-level-pixel-label)
+  * [技巧](#技巧)
+    * [空洞卷积](#空洞卷积)
+    * [ASPP](#aspp)
+    * [全连接CRF](#全连接crf)
+  * [实验](#实验)
+    * [失败的模式](#失败的模式)
+  * [结论](#结论)
+  * [参考](#参考)
+
 ## 摘要
 
 在这项工作中, 我们用深度学习来解决语义图像分割的任务, 并做出三个主要贡献, 这些贡献在实验上显示具有实质性的实用价值.

diff --git a/...ge-Aware Convolution Neural Network Based Salient Object Detection论文速览(2019).md b/...ge-Aware Convolution Neural Network Based Salient Object Detection论文速览(2019).md
@@ -2,6 +2,15 @@
 
 > Wenlong Guan, Tiantian Wang, Jinqing Qi, Lihe Zhang and Huchuan Lu, Edge-Aware Convolution Neural Network Based Salient Object Detection, IEEE Signal Processing Letters, Vol. 26, No. 1, P114-118,2019 [PDF(baidu)](https://pan.baidu.com/s/1RsgfISTe7MHprUtwtvYbvQ) [PDF(google)](https://drive.google.com/file/d/1vF4HqiKE7iapWPPmzvttbrz68PEspMuB/view?usp=sharing)
 
+* [Edge-Aware Convolution Neural Network Based Salient Object Detection](#edge-aware-convolution-neural-network-based-salient-object-detection)
+  * [概要](#概要)
+  * [架构解释](#架构解释)
+    * [整体结构](#整体结构)
+    * [解码器结构](#解码器结构)
+    * [测试结果](#测试结果)
+      * [量化比较](#量化比较)
+      * [消融实验](#消融实验)
+
 ## 概要
 
 近年来，目标检测受到了广泛的关注。在这篇论文中，提出了一种新颖的目标检测算法，该算法将全局信息信息与低级边缘特征相结合。

diff --git a/图像分割/Pyramid Scene Parsing Network翻译(2016).md b/图像分割/Pyramid Scene Parsing Network翻译(2016).md
@@ -1,5 +1,21 @@
 # Pyramid Scene Parsing Network(2016)
 
+* [Pyramid Scene Parsing Network(2016)](#pyramid-scene-parsing-network2016)
+  * [概要](#概要)
+  * [引言](#引言)
+  * [相关工作](#相关工作)
+  * [Pyramid Scene Parsing Network](#pyramid-scene-parsing-network)
+    * [重要观察](#重要观察)
+      * [匹配关系](#匹配关系)
+      * [混淆类别](#混淆类别)
+      * [不显眼的类别](#不显眼的类别)
+      * [观察总结](#观察总结)
+    * [Pyramid Pooling Module](#pyramid-pooling-module)
+    * [Network Architecture](#network-architecture)
+  * [Deep Supervision for ResNet-Based FCN](#deep-supervision-for-resnet-based-fcn)
+  * [实验](#实验)
+  * [总结](#总结)
+
 ## 概要
 
 场景解析对于不受限制的开放式场景和多样化场景具有挑战性. 在本文中, 通过金字塔池化模块以及提出的金字塔场景解析网络（PSPNet）实现的, 通过基于不同区域的上下文聚合来利用全局上下文信息.

diff --git a/...i-Path Refinement Networks for High-Resolution Semantic Segmentation翻译(2016).md b/...i-Path Refinement Networks for High-Resolution Semantic Segmentation翻译(2016).md
@@ -1,5 +1,22 @@
 # RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation(2016)
 
+* [RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation(2016)](#refinenet-multi-path-refinement-networks-for-high-resolution-semantic-segmentation2016)
+  * [概要](#概要)
+  * [引言](#引言)
+  * [相关工作](#相关工作)
+  * [背景](#背景)
+  * [提出的方法](#提出的方法)
+    * [Multi-Path Refinement](#multi-path-refinement)
+    * [RefineNet](#refinenet)
+      * [Residual convolution unit(RCU)](#residual-convolution-unitrcu)
+      * [Multi-resolution fusion](#multi-resolution-fusion)
+      * [Chained residual pooling](#chained-residual-pooling)
+      * [Output convolutions](#output-convolutions)
+    * [Identity Mappings in RefineNet](#identity-mappings-in-refinenet)
+  * [实验](#实验)
+  * [改进版本](#改进版本)
+  * [结论](#结论)
+
 ## 概要
 
 最近, 很深的卷积神经网络(CNN)在目标识别中有着出色的表现, 也成为了密集分类问题的首选, 例如语义分割等问题. 然而, 在CNN中重复的下采样操作, 像池化或卷积, 导致显著降低了初始图像的分辨率. 这里, 论文提出的RefineNet, 一个通用的多路径优化网络, 明确地沿着下采样流程, 利用所有可用的信息, 使高分辨率预测使用远距离残余链接. 这样, 捕获高级语义特征的更深的层, 可以直接提炼使用来自早期卷积的细粒度特征. RefineNet的单个组件使用残余链接, 接在恒等映射之后, 它允许有效的端到端训练. 进一步, 论文引入了链式残差池化(chained residual pooling), 来以一种有效方式捕捉丰富背景内容.

diff --git a/...rsing via Dense Recurrent Neural Networks with Attentional Selection总结(2018).md b/...rsing via Dense Recurrent Neural Networks with Attentional Selection总结(2018).md
@@ -1,5 +1,22 @@
 # Scene Parsing via Dense Recurrent Neural Networks with Attentional Selection(2018)
 
+* [Scene Parsing via Dense Recurrent Neural Networks with Attentional Selection(2018)](#scene-parsing-via-dense-recurrent-neural-networks-with-attentional-selection2018)
+  * [文章要点](#文章要点)
+    * [计算流程改进](#计算流程改进)
+      * [相关参数的定义](#相关参数的定义)
+      * [原本密集RNN的计算流程](#原本密集rnn的计算流程)
+      * [考虑注意力模型](#考虑注意力模型)
+      * [集成到整体后的密集RNN](#集成到整体后的密集rnn)
+  * [实际效果](#实际效果)
+    * [基线比较](#基线比较)
+    * [在PASCAL Context上的结果比较](#在pascal-context上的结果比较)
+    * [在MIT ADE20K上的结果比较](#在mit-ade20k上的结果比较)
+    * [对于Cityscopes上的结果比较](#对于cityscopes上的结果比较)
+  * [消融实验](#消融实验)
+  * [模型复杂度研究](#模型复杂度研究)
+  * [总结](#总结)
+  * [参考](#参考)
+
 ## 文章要点
 
 1. 使用无向有环图对图像进行建模, 并使用密集连接改进

diff --git a/...ep Convolutional Encoder-Decoder Architecture for Image Segmentation翻译(2015).md b/...ep Convolutional Encoder-Decoder Architecture for Image Segmentation翻译(2015).md
@@ -1,5 +1,16 @@
 # SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation(2015)
 
+* [SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation(2015)](#segnet-a-deep-convolutional-encoder-decoder-architecture-for-image-segmentation2015)
+  * [结构](#结构)
+  * [训练](#训练)
+    * [数据集](#数据集)
+    * [初始化](#初始化)
+    * [损失](#损失)
+    * [指标](#指标)
+  * [测试](#测试)
+  * [想法](#想法)
+  * [参考](#参考)
+
 ## 结构
 
 ![1545042513014](assets/1545042513014.png)

diff --git a/图像分割/Semantic Human Matting翻译(2018).md b/图像分割/Semantic Human Matting翻译(2018).md
@@ -2,6 +2,39 @@
 
 > <http://arxiv.org/abs/1809.01354v1>
 
+* [Semantic Human Matting](#semantic-human-matting)
+  * [Abstract](#abstract)
+  * [Introduction](#introduction)
+  * [Related works](#related-works)
+    * [semantic segmentation](#semantic-segmentation)
+    * [image matting methods](#image-matting-methods)
+    * [automatic matting system](#automatic-matting-system)
+  * [Human matting dataset](#human-matting-dataset)
+  * [Our method](#our-method)
+    * [Trimap generation: T-Net](#trimap-generation-t-net)
+    * [Matting network: M-Net](#matting-network-m-net)
+    * [Fusion Module](#fusion-module)
+    * [Loss](#loss)
+    * [Implementation Detail](#implementation-detail)
+      * [T-Net pre-train](#t-net-pre-train)
+      * [M-Net pre-train](#m-net-pre-train)
+      * [End-to-end training](#end-to-end-training)
+      * [Testing](#testing)
+  * [Experiments](#experiments)
+    * [Experimental Setup](#experimental-setup)
+      * [Dataset](#dataset)
+      * [Measurement](#measurement)
+      * [Baselines](#baselines)
+    * [Performance Comparison](#performance-comparison)
+    * [Automatic Method vs. Interactive Methods](#automatic-method-vs-interactive-methods)
+    * [Evaluation and Analysis of Different Components](#evaluation-and-analysis-of-different-components)
+      * [The Effect of End-to-end Training](#the-effect-of-end-to-end-training)
+      * [The Evaluation of Fusion Module](#the-evaluation-of-fusion-module)
+      * [The Effect of Constraint Lt](#the-effect-of-constraint-lt)
+      * [Visualization of Intermediate Results](#visualization-of-intermediate-results)
+    * [Applying to real images](#applying-to-real-images)
+  * [Conclusion](#conclusion)
+
 ## Abstract
 
 Human matting, high quality extraction of humans from natural images, is crucial for a wide variety of applications. Since the matting problem is severely under-constrained, most previous methods require user interactions to take user designated trimaps or scribbles as constraints. This user-in-the-loop nature makes them difficult to be applied to large scale data or time-sensitive scenarios. In this paper, instead of using explicit user input constraints, we employ implicit semantic constraints learned from data and propose an automatic human matting algorithm Semantic Human Matting (SHM).

diff --git a/图像分割/U-Net: Convolutional Networks for Biomedical Image Segmentation翻译(2015).md b/图像分割/U-Net: Convolutional Networks for Biomedical Image Segmentation翻译(2015).md
@@ -1,5 +1,17 @@
 # U-Net: Convolutional Networks for Biomedical Image Segmentation(2015)
 
+* [U-Net: Convolutional Networks for Biomedical Image Segmentation(2015)](#u-net-convolutional-networks-for-biomedical-image-segmentation2015)
+  * [关键点](#关键点)
+  * [网络结构流程](#网络结构流程)
+  * [训练](#训练)
+  * [一些疑惑](#一些疑惑)
+  * [一些想法](#一些想法)
+    * [运算](#运算)
+    * [组合](#组合)
+    * [连接](#连接)
+    * [拼接](#拼接)
+  * [参考链接](#参考链接)
+
 ## 关键点
 
 * 编码解码结构

diff --git a/图像分割/UNet++: A Nested U-Net Architecture for Medical Image Segmentation简介(2018).md b/图像分割/UNet++: A Nested U-Net Architecture for Medical Image Segmentation简介(2018).md
@@ -4,6 +4,12 @@
 >
 > https://zhuanlan.zhihu.com/p/44958351
 
+* [UNet++: A Nested U-Net Architecture for Medical Image Segmentation(2018)](#unet-a-nested-u-net-architecture-for-medical-image-segmentation2018)
+  * [结构](#结构)
+  * [关于深度的思考](#关于深度的思考)
+  * [使用多尺度的特征](#使用多尺度的特征)
+  * [深监督与剪枝](#深监督与剪枝)
+
 ## 结构
 
 文章对U-Net改进的点主要是skip connection.

diff --git a/图像分割/条件随机场CRF总结.md b/图像分割/条件随机场CRF总结.md
@@ -2,6 +2,15 @@
 
 > [条件随机场CRF(一)从随机场到线性链条件随机场](https://www.cnblogs.com/pinard/p/7048333.html#undefined)
 
+* [条件随机场](#条件随机场)
+  * [简介](#简介)
+  * [数学描述](#数学描述)
+  * [参数化](#参数化)
+  * [向量化](#向量化)
+  * [矩阵化](#矩阵化)
+  * [基本问题](#基本问题)
+  * [应用](#应用)
+
 ## 简介
 
 条件随机场（conditional random field，简称 CRF），是一种鉴别式机率模型，是随机场的一种，常用于标注或分析序列资料，如自然语言文字或是生物序列。 如同马尔可夫随机场，条件随机场为无向性之图模型，图中的顶点代表随机变量，顶点间的连线代表随机变量间的相依关系，在条件随机场当中，随机变量 Y 的分布为条件机率，给定的观察值则为随机变量 X。原则上，条件随机场的图模型布局是可以任意给定的，一般常用的布局是链接式的架构，链接式架构不论在训练（training）、推论（inference）、或是解码（decoding）上，都存在有效率的算法可供演算。