🧠
AI
  • Artificial Intelligence
  • Intuitive Maths behind AI
    • Probability
    • Information Theory
    • Linear Algebra
    • Calculus
  • Overview
  • Research Ideas and Philosophy
  • Basic Principles
  • Information Theory
    • Entropy
    • Log Probability
  • Probability & Statistics
    • Random Variables
    • Probability
      • Probablistic Equations
      • Bayes Theorem
      • Probability Distributions & Processes
    • Statistics
      • Measures
      • Z-Scores
      • Covariance and Correlation
      • Correlation vs Dependance
    • Mahalanobis vs Chi-Squared
    • Uncertainty
    • Statistical Inference
      • Graphical Models
      • Estimator vs Parameter
      • Estimation
      • Bayesian/Probabilistic Inference
        • Probabilistic Modelling
        • Problems of Bayesian Inference
        • Conjugate Priors
        • Dirichlet Distribution/Process
        • Posterior Predictive Distribution
      • Sampling-Based Inference
    • Sampling
      • Rejection Sampling
      • Reservoir Sampling
      • Thompson Sampling
    • Bayesian Inference
    • Regression
    • Markov
    • Monte Carlo
      • Monte Carlo Estimators
      • Importance Sampling
    • Kernel Density Estimation
    • Gaussian Processes
    • Gaussian Soap Bubble
  • Linear Algebra
    • Vector Space and Matrices
    • Geometry of System of Linear Equations
    • Determinants
    • Transformations
    • Geometrical Representation
    • Positive (Semi)Definite Matrices
    • Matrix Interpretation
    • Dot Product as Linear Transformation and Duality of Vector-Linear Transformation
    • Norms
    • Linear Least Square
    • Matrix Decomposition
      • QR Decomposition
      • Cholesky Decomposition
      • Eigen Value Decomposition
      • SVD - Singular Value Decomposition
    • Matrix Inversion
    • Matrix Calculus
    • Matrix Cookbook
    • Distributed Matrix Algebra
    • High Dimensional Spaces
  • Optimization
    • Derivatives
      • Partial Derivative
      • Directional Derivative
      • Gradient
      • Jacobian
    • Regularization
    • Gradient Descent
    • Newton's Method
    • Gauss-Newton
    • Levenberg–Marquardt
    • Conjugate Gradient
    • Implicit Function Theorem for optimization
    • Lagrange Multiplier
    • Powell's dog leg
    • Laplace Approximation
    • Cross Entropy Method
    • Implicit Function Theorem
  • Statistical Learning Theory
    • Expectation Maximization
  • Machine Learning
    • Clustering
    • Bias Variance Trade-off
  • Deep Learning
    • PreProcessing
    • Convolution Arithmetic
    • Regularization
    • Optimizers
    • Loss function
    • Activation Functions
    • Automatic Differentiation
    • Softmax Classifier and Cross Entropy
    • Normalization
    • Batch Normalization
    • Variational Inference
    • VAE: Variational Auto-Encoders
    • Generative vs Discriminative
      • Generative Modelling
    • Making GANs train
    • Dimensionality of Layer Vs Number of Layers
    • Deep learning techniques
    • Dilated Convolutions
    • Non-Maximum Suppression
    • Hard Negative Mining
    • Mean Average Precision
    • Fine Tuning or Transfer Learning
    • Hyper-parameter Tuning
  • Bayesian Deep Learning
    • Probabilistic View
    • Uncertainty
    • Variational Inference for Bayesian Neural Network
  • Reinforcement Learning
    • General
    • Multi-armed Bandit
    • Imitation Learning
    • MDP Equations
    • Solving MDP with known Model
    • Value Iteration
    • Model Free Prediction and Control
    • Off Policy vs On Policy
    • Control & Planning from RL perspective
    • Deep Reinforcement Learning
      • Value Function Approximation
      • Policy Gradient
        • Algorithms
    • Multi Agent Reinforcement Learning
    • Reinforcement Learning - Sutton and Barto
      • Chapter 3: Finite Markov Decision Processes
      • Chapter 4: Dynamic Programming
    • MBRL
  • Transformers
    • Tokenziation
    • Embedding
      • Word Embedding
      • Positional Encoding
    • Encoder
    • Decoder
    • Multi-head Attention Block
    • Time Complexities of Self-Attention
    • KV Cache
    • Multi-head Latent Attention
    • Speculative Decoding
    • Flash Attention
    • Metrics
  • LLMs
    • LLM Techniques
    • LLM Post-training
    • Inference/Test Time Scaling
    • Reasoning Models
    • Reward Hacking
  • Diffusion Models
    • ImageGen
  • Distributed Training
  • State Space Models
  • RLHF
  • Robotics
    • Kalman Filter
    • Unscented Kalman Filter
  • Game Theory and ML
    • 1st Lecture - 19/01
    • Lecture 2 - 22/01
    • Lecture 4: Optimization
  • Continual Learning
    • Lecture - 21/01
    • iCaRL: Incremental Classifier and Representation Learning
    • Variational Continual Learning
  • Computer Vision
    • Hough Transform
    • Projective Geometry
      • Extrinsic and Intrinsic Parameters
      • Image Rectification
    • Tracking
    • Optical Flow
    • Harris Corner
    • Others
  • Papers
    • To Be Read
    • Probabilistic Object Detection and Uncertainty Estimation
      • BayesOD
      • Leveraging Heteroscedastic Aleatoric Uncertainties for Robust Real-Time LiDAR 3D Object Detection
      • Gaussian YOLOv3
      • Dropout Sampling for Robust Object Detection in Open-Set Condition
      • *Sampling Free Epistemic Uncertainty Estimation using Approximated Variance Propagation
      • Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics
      • Can We Trust You? On Calibration of Probabilistic Object Detector for Autonomous Driving
    • Object Detection
    • Temporal Fusion in Object Detection/ Video Object Detection
    • An intriguing failing of convolutional neural networks and the CoordConv solution
    • A Neural Algorithm of Artistic Style - A.Gatys
  • Deep Learning Book
    • Chapter 4: Optimization
    • Chapter 5: Machine Learning Basics
    • Chapter 6: Deep FeedForward Networks
  • Python
    • Decorators
    • Packages
      • Pip
    • Gotchas
    • Async functions
  • Computer Science
  • TensorFlow
  • Pytorch
    • RNN/LSTM in Pytorch
    • Dataset/ Data loader
    • Resuming/Loading Saved model
  • Programming
    • Unit Testing
    • How to write code
  • General Software Engineering
    • SSH tunneling and Ngrok
  • How To Do Research
  • Resources
  • ROS for python3
  • Kitti
Powered by GitBook
On this page
  • Camera Model and Parameters
  • Homogenous Representation
  • Homography
  • Matrix
  • Camera Geometry/Single View Geometry
  • Camera Model
  • Projection Matrix: From 3D to image coordinates
  1. Computer Vision

Projective Geometry

PreviousHough TransformNextExtrinsic and Intrinsic Parameters

Last updated 5 years ago

Camera Model and Parameters

  • World Frame: a fixed coordinate system for representing objects (points, lines, surfaces, etc.) in the world.

  • Camera Frame: coordinate system that uses the camera center as its origin (and the optic axis as the Z-axis)

  • Image or retinal plane: plane on which the image is formed, note that the image plane is measured in camera frame coordinates (mm)

  • Image Frame: coordinate system that measures pixel locations in the image plane.

  • Intrinsic Parameters: Camera parameters that are internal and fixed to a particular camera/digitization setup. Allow a mapping between camera coordinates and pixel coordinates in the image frame`.

  • Extrinsic Parameters: Camera parameters that are external to the camera and may change with respect to the world frame. Define the location and orientation of the camera with respect to the world frame.

Homogenous Representation

Each point (x,y) in the image is actually a ray in 3D world, the ray can be reresented as (kx, ky,k) and when k=1 it gives the point in the image. Get it like this, all the point on the ray in the 3D world will be at the same point on image. Also, that ray will through the center of the camera. Hence homogenous representation is basically way to denote the 2D point in image.

Example:

Lets consider three points x=[6,9,3]x=[6,9,3]x=[6,9,3] , y=[12,18,6]y=[12,18,6]y=[12,18,6] and z=[2,3,1] in 3D world, these 3 points lie on the same ray passing through the origin. So, basically all these three point will be at map to same point in the image plane. And the point zzz is the point which lie exactly at the image plane. Hence so given a point (x,y) (x,y)(x,y) in the image place, it can map to any point (kx,ky,k)∀k(kx, ky,k) \forall k(kx,ky,k)∀k and this is called projective space.

Homography

Homography is the condition when image of same plane is taken with two different position. There is a mapping between two points xxx and x′x'x′ in both images respectively which represent the same point XXX in the 3D world. This mapping between coordinates of the image of same plane is called Homography.

Matrix

  • The homography preserves the colliniearity of points. i.e. if three points are collinear in one image then they will be collinear even after the homography.

  • Since HHH and kHkHkH will be same here, hence numper of unknown in HHH are 8 and not 9.

  • Since 8 unknowns, hence we will need 4 corresponding points to find the HHH

Camera Geometry/Single View Geometry

Camera Model

Projection Matrix: From 3D to image coordinates

  • This is important as it can be used to find relation between 3D world and 2D image.

  • Depth From Stereo For these we need following information:-

    • Info about relative position and orientation of two cameras wrt each other.

    • Point correspondence in two images from camera.

  • Lectures 200-205 are important.

  • Lecture 219-227 are for extrinsic geometry of camera(Camera-world transform.)

  • Essential Matrix is a geometric constraint between world coordinate of point in image taken by different camera position. X`EX = 0. X` and X are vector of point in two camera coordinate system. The E matrix is composed of Rotation and Translation Matrix between two camera coordinate system.

Fundamental Matrix defines the relation of a point in two image coordinates. P`FP=0, Where P and P` are coordinates of a point in two images.Fundamental Matrix is composed of Extrinsic & Intrinsic Parameters. F matrix is always singular. Hence F matrix is sufficient to find corresponding point in two images taken from different camera position and from different camera intrinsics. Also Homographies are also used as relation between two planes of images. Therefore, fundamental matrix and homography matrix has some relation between them.

Here x,y,z are coordinate in 3D world and u,v are coordinate of those points in image.

Therefor all pairs of parallel lines in 3D world meet at a point in image.

Intrinsic Matrix:- (u,v) are pixel coordinate in image.