Computer Vision

[목차]

• Preliminary
• Image Classification
  • Image Processing Introduction
  • Image Classification + CV
  • Image Classification + DL
• CNN Architectures
• Object Detection
• Semantic Segmentation
  • Instance-aware Semantic Segmentation
• Metric Learning
• Video Vision
  • Video Classification + CV
  • Video Classification + DL
  • Visual Tracking
• Model Fitting
• Camera Models

Preliminary

• Linear Algebra
• Bayes’ Theorem

---

Image Classification

Image Processing Introduction

• Convolution
  • smoothing
    • Bilinear
    • Average
    • Gaussian
• Edge
  • Gradient Image
• Corners: Harris Corner Detection
  • Eigen Decomposition
• Blob
  • Laplace of Gaussian; LoG

Image Classification + CV

• SIFT (2004)

1. Finding Scale-Space Extrema
2. Keypoint Filtering
3. Orientation Assignment
4. Calculating Descriptor

• Spatial Pytramid Matching (2006)
• Discrimative vs. Geneartive Model

Image Classification + DL

• MLP
• Loss Functions
• Gradient Descent
  • SGD
  • Momentum
• CNN
• Overfitting Issue
  • Drop out
  • Weight decay
  • Early Stopping
  • Network Initialization
    • Learning from scratch
    • Xavier Initialization (2010)
    • He Initialization (2015)

---

CNN Architectures

• LeNet (1998)
• AlexNet
  • LRN; Local response Normalizatioin
• VGGNet (2014)
• ResNet (2016)
  • Degrading Problem
  • Skip Connection
  • Batch Normalization (2015)
• Beyond ResNet
  • DenseNet (2017)
    • Channel-wise concatenation
  • SENet (2017)
    • Squeeze & Excitation

---

Object Detection

• Support Vector Machine
  • Linear SVM + Separable Case
  • Linear SVM + Non-Separable Case
    • Soft margin
  • Non-Linear SVM
    • Kernel Method
  • Multi-Class SVM
• Pedestrian Detection + SVM (2005)
  • HOG Histogram of Orientated Gradient+ SVM

• R-CNN Region-base CNN (2014)
  • Object proposal
    • Selective Search
• Fast R-CNN (2015)
  • ROI pooling
• Faster R-CNN (2015)
  • Fast R-CNN + RPN Region Proposal Network

---

Semantic Segmentation

• Fully Convolutional Network (FCN Family)
  • FCN (2015)
  • DeepLab
• Convolutional Encoder-Decoders
  • U-Net
  • DeConvNet (2015)

• FCN (2015)
  • 1x1 conv
  • adding skip connection
• DeepLab (2017)
  • Atrous Convolution
  • CRF; Fully-Connected Conditional Random Field
• Pyramid Scene Parsing Network (2017)
  • Pyramid pooling module
• Context Encoding Network (2018)
  • Attention module

• DeConvNet (2015)
  • conv - deconv
  • pooling - unpooling

Instance-aware Semantic Segmentation

• Multi-task Network Cascades (2016)
• Multi-scale Patch Aggregation (2016)
• Mask R-CNN (2017)
  • ROI Align

---

Metric Learning

• Pairwise & Triplet Metric
• Mahalanobis Distance
• A first approach to distance metric learning (Pairwise)
• Large Margin Nearest Neighbor(LMNN) (Triplet)
• Metric Learning + DL

---

Video Vision

Video Classification + CV

• Optical Flow
  • (가정) Color constancy
  • (가정) Small motion
  • Lukas-Kanade Flow
• STIP; Space-Time Interest Point (2005)
• Dense Trajectory

Video Classification + DL

• 3D CNN (2010)
• C3D (2015)
• Time Information Fusion (2014)
  • Sing Frame
  • Late Fusion
  • Early Fusion
  • Slow Fusion
• Two-Stream Cconvolutional Network (2014)

---

Visual Tracking

• Probabilistic Tracking
• Sequential Density Estimation
• Kalman Filter
• Particle Filtering

---

Model Fitting

• Least Square
  • Ordinary Linear Least Square
  • Total Linear Least Square
• RANSAC RANdom SAmple Consensus
• Hough Transform

---

Camera Models

• 2D Objects
• 2D Transformations
  • Translation
  • Euclidean transform
  • Similarity transform
  • Affine transform
  • Projective transform
• 3D Objects
  • homogeneous coordinates; $\overline{x} = [x, y, z, 1]$
• Pinhole Model
  • Intrinsic Parameters
  • Extrinsic Parameters
• Camera Clibration
  • Estimate camera parameters matrix