Computer Vision

Computer Vision

About This Series

Computer vision is the field that estimates the shape, position, category, motion and three-dimensional structure of objects from visual information such as images and video, in order to understand a scene. This series is organised around geometric computer vision: it starts with the mathematical models of cameras and progresses through feature detection, stereo vision, Structure from Motion, and on to SLAM and NeRF.

Computer vision is widely used in autonomous driving, AR/VR, robotics, medical imaging, and many other fields.

Learning by Level

Introduction

Image and Camera Fundamentals

  • Digital Image Fundamentals
  • Color Spaces
  • Pinhole Camera Model
  • Homogeneous Coordinates and Transformation Matrices
  • Camera Matrix
  • History of Computer Vision (reading)

Basic

Features and Calibration

  • Feature Detection and Description (Harris, SIFT, ORB)
  • Feature Matching
  • Robust Estimation and RANSAC
  • Camera Calibration

Intermediate

Stereo and 3D Reconstruction (in preparation)

  • Epipolar Geometry
  • Stereo Vision
  • Structure from Motion
  • Multi-View Stereo

Advanced

SLAM and Cutting-Edge Methods (in preparation)

  • Visual Odometry
  • SLAM
  • Deep Learning-Based Methods
  • NeRF / 3D Gaussian Splatting (scene representation and novel-view synthesis)

Learning Path

Introduction Camera Basics Basic Features & Calib. Intermediate 3D Reconstruction Advanced SLAM & NeRF Intro: Images, camera models, coordinate transforms, projection Basic: Features, matching, RANSAC, calibration Intermediate: Epipolar geometry, stereo, SfM, MVS Advanced: VO, SLAM, deep learning, NeRF, 3DGS Geometric Computer Vision: From 2D Images to the 3D World 2D Image Pixel Coordinates (u, v) Feature Matching Multiple Views Epipolar Geometry Triangulation 3D Recovery 3D Reconstruction Poses / Point Cloud (X, Y, Z)

* 3D recovery needs camera information and enough parallax between viewpoints; images alone may not fix the real-world scale.

Key Topics

Camera Geometry

Pinhole model, perspective projection, intrinsic and extrinsic parameters.

Feature Detection and Description

Corner detection such as Harris (choosing points), and methods such as SIFT and ORB that both detect points and encode their surroundings as descriptors.

3D Reconstruction

Depth from stereo vision, camera poses and sparse 3D points from SfM, and dense shape from MVS.

SLAM

Simultaneous localization and mapping: estimating position while building a map of the environment. This series covers Visual SLAM, which uses images (SLAM in general also uses sensors such as LiDAR).