Computer Vision -- CS 570, Department of Computer Science, PU

CS570 Computer Vision
Nazar Khan

Human beings (and even animals) "look" at the real-world and extract extremely accurate information extremely efficiently. Computers can fail catastrophically at this task! In this course we look into why "Vision" is a difficult problem to solve and we go through successful, mathematically well-founded techniques used to solve the Vision problem.

This course is a useful application of mathematical concepts from Linear Algebra and Calculus. Therefore, the students could do well by brushing up on their Linear Algebra, Calculus and programming skills before taking this class. The techniques learned here can be useful for other areas such as Image Processing, Machine Learning, Artificial Intelligence and Computer Graphics.

CS 570 is a graduate course worth 3 credit hours.

Lectures: Monday and Wednesday, 10:15 a.m. - 11:45 a.m. Room 14, FCIT, Allama Iqbal (Old) Campus
Google Classroom:
Office Hours: Monday, 2:30 p.m. - 3:30 p.m. or by appointment

Grading:

Assignments	35%
Project	15%
Quizzes	5%
Tests	5%
Mid-Term	15%
Final	25%

Prerequisites

Ability to code
Basic Calculus (Differentiation, Partial derivatives, Chain rule)
Linear Algebra (Vectors, Matrices, Dot-product, Orthogonality, Eigenvectors)

Books and Other Resources

No single book will be followed as the primary text. Helpful online and offline resources include:

Image Processing

Digital Image Processing by R. C. Gonzalez, R. E. Woods.

Computer Vision

Introductory techniques for 3D Computer Vision by Trucco and Verri
Image Processing, Analysis, and Machine Vision by Sonka, Hlavac, Roger
Computer Vision: Algorithms and Applications by Richard Szeliski (http://szeliski.org/Book/)
Multiple View Geometry in Computer Vision by Hartley and Zisserman
Computer Vision: A Modern Approach by Forsyth and Ponce
Computer Vision by Linda G. Shapiro, George C. Stockman

Lectures

#	Topics	Slides	Videos	Recitations	Readings	Miscellaneous
1	Introduction What is Computer Vision? Computer Vision vs. Biological Vision -- The Grand Deception! Applications of Computer Vision Course Details	Introduction to Computer Vision	Video
2	Background Math I Notation Inner Product Euclidean Norm Outer Product Matrix and Vector Calculus	Background Mathematics	Video	Friday November 26: Recitation 1 Python Numpy Video Recitation 1		Take-home Quiz 1
3	Background Math II Matrices as Linear Operators Eigenvectors Optimisation Constrained Optimisation Taylor Series Approximation Cartesian vs. Image axis	Background Mathematics	Video
4	Image Filtering Convolution Properties of convolution Mean Filtering Gaussian Filtering Non-linear filtering cannot be performed via convolution	Image Filtering	Video	Friday December 3: Recitation 2 Working with images Basic image handling using NumPy PIL OpenCV Filtering Point operations Image Thresholding Neighborhood operations Convolution Averaging Filter Gaussian Blurring Median Blurring Bilateral Filtering Histogram equalization Video Recitation 2		Take-home Quiz 2
5	Derivative Approximations Gradient magnitude and angle via partial derivatives atan vs. atan2 Derivative approximations via Taylor's formula 1st-derivative using central difference 2nd-derivative using central difference 1st-derivative using forward difference 1st-derivative using backward difference	Derivative Approximations	Video
6	Derivative Filtering & Edge Detection Convolution masks for derivative filtering Smoothing before computing derivatives Forward, Backward and Central Difference Derivative of Gaussian Filter Sobel operator Naive Edge Detection The Canny Edge Detector Gradient computation Non-maxima suppression Hysteresis thresholding	Derivative Filtering & Edge Detection	Video	Saturday December 11: Recitation 3 Edge Detection Sobel Filter Prewitt’s Filter Laplacian Filter Canny Edge Detector Video Processing Applications of Edge Detection Road Lane Detection Edge-based Face Recognition Video Recitation 3		Take-home Quiz 3 Assignment 1
7	The Structure Tensor Comparing patches -- SSD Quadratic form for SSD Taylor's approximation in SSD Weighted SSD via Gaussian convolution Structure Tensor -- Geometry and Algebra	Structure Tensor	Video
8	Corner Detection Harris Corner Detector Rohr Corner Detector Scale Space and Gaussian Pyramids	Corner Detection	Video	Saturday December 18: Recitation 4 Corner Detection Harris corner detector Shi-Tomasi corner detector Scale-space via Gaussian pyramid Video Recitation 4		Take-home Quiz 4 Assignment 2
9	Local Image Descriptors I Sum-of-Squared-Differences (SSD) Normalized Cross Correlation (NCC) SIFT Keypoint Detection	Local Image Descriptors	Video
10	Local Image Descriptors II SIFT Descriptor Matching SIFT Descriptors	Local Image Descriptors	Video	Friday December 24: Recitation 5 Feature Extraction SIFT (Scale Invariant Feature Transform) SURF (Speeded-Up Robust Features) ORB (Oriented FAST and Rotated Brief) Feature Matching SIFT robustness to Rotation, Scale, and Blurriness Object Matching in Video Video Recitation 5		Take-home Quiz 5
11	Hough Transform Slope-intercept representation of lines Polar representation of lines Line detection Circle detection Speed-up via gradient information	Hough Transform	Video
12	Deep Learning -- I Deep Learning is unavoidable Machine Learning Deep Learning The Artificial Neuron Neural Networks Loss Functions	Deep Learning	Video	Friday January 14: Recitation 6 Hough Transform Line Detection Circle Detection Applications Red Blood Cells Estimation Iris Detection Traffic Signs Detection Video Recitation 6		Take-home Quiz 6 Assignment 3 Project Proposal
13	Deep Learning -- II Multiclass Classification Activation Functions Regularization	Deep Learning	Video
14	Convolutional Neural Network (CNN) Convolutional Filters Subsampling Fully Connected Layers 1x1 Convolutions Depthwise separable Convolutions Transposed Convolutions Unpooling Fully Convolutional Networks Semantic Segmentation Residual Blocks	Convolutional Neural Network (CNN)	Video	Friday January 21: Recitation 7 Keras basics About train, val and test sets Building blocks in Keras Classifying cats vs dogs using CNN Video Recitation 7
	Mid-Term Exam
15	Object Detection, Classification and Segmentation via Mask R-CNN Detection vs. Classification vs. Segmentation Overall Architecture	Object Detection, Classification and Segmentation via Mask R-CNN	Video
16	Object Detection, Classification and Segmentation via Mask R-CNN Feature Pyramid Network (FPN) Region Proposal Network (RPN) RoIAlign Classification Bounding Box Regression Instance Segmentation	Object Detection, Classification and Segmentation via Mask R-CNN	Video	Friday February 4: Recitation 8 Mask R-CNN Custom Dataset Annotation Training Video Recitation 8
17	2D Spatial Transformations Matrix ≡ Linear Transformation Scaling, Shear, Rotation Translation is not linear in ℝ² Homogenous Coordinates make translation linear in ℙ² 2D Affine Transformation (Scaling, Shear, Rotation, Translation) 6 degrees of freedom 2D Projective Transformation (Homography) 8 degrees of freedom	2D Spatial Transformations	Video
18	Estimating and Applying Transformations Recovering best affine transformation from correspondences Recovering best projective transformation from correspondences -- Direct Linear Transform (DLT)	Estimating Transformations	Video	Friday February 11: Recitation 9 Affine Transformation Scaling Shearing Translation Rotation Estimation of affine transform Projective transformation (Homography) Perspective correction Virtual Billboard Add sponsors to sponsor-board in a video Video Recitation 9		Take-home Quiz 7
19	... DLT continued Image Warping Problems with naïve warping Inverse transformation Bilinear Interpolation	Image Warping	Video
20	Robust Estimation Outliers Least-squares models are sensitive to outliers Robust Estimation via RANSAC Line Fitting Homography Estimation	RANSAC	Video	Friday February 18: Recitation 10 Robust line fitting using RANSAC Object tracking in videos Image stitching Video Recitation 10
21	Optic Flow -- Local I Gray value constancy Linearized Optic Flow Constraint (OFC) Aperture Problem Normal Flow Local Method of Lucas & Kanade	Optic Flow -- Local	Video
22	Optic Flow -- Local II Local Method of Lucas & Kanade Flow Classification via Structure Tensor Flow Visualization	Optic Flow -- Local	Video	Friday February 25: Recitation 11 Optic Flow in OpenCV Motion Estimation and Tracking Video Recitation 11		Take-home Quiz 8 Assignment 4
23	Optic Flow -- Global Global Method of Horn & Schunck Data Term Smoothness Term Regularization Parameter Functions versus Functionals Calculus of Variations Euler-Lagrange Equations Fixed-point Iterations	Optic Flow -- Global	Video			Take-home Quiz 9
24	Camera Geometry Camera Obscura Pinhole Camera Model Camera Matrix Intrinsic Parameters Projection Extrinsic Parameters	Camera Geometry	Video
25	Camera Anatomy Camera Center Why do parallel lines meet in images? Points at infinity Vanishing points What do the columns of P tell us? What do the rows of P tell us?	Camera Anatomy	Video	Friday March 11: Recitation 12 Camera Calibration in OpenCV Pose estimation and augmented reality Video Recitation 12
26	Conclusion What was covered? What were the general principles? What was not covered?	Conclusion
	Final Exam