Introduction to Computer Vision - Lecture 1 Notes

Welcome and Course Overview

  • Goal: Convert light into meaning (geometric, semantic).

  • Computer Vision: Transforming images to useful information.

  • Key questions: Where, what objects, how many, who, camera location, object positions, actions?

  • Themes: What questions can be answered using images, and what methods can be used to find these answers?

"Useful Information"

  • Images are rich in information.

  • A tiny image of 10x10 pixels with grayscale values from 0 to 255 can have 255100255^{100} different representations.

  • Feature extraction is an intermediate step, not the end goal.

Applications of Computer Vision

  • Robotics, medical applications, 3D modeling, driving, mobile devices, accessibility.

  • Example: OCR/document verification requiring alignment and feature matching.

  • Homography: Any two images of the same planar surface are related by a homography.

    • p<em>i′∝K</em>1H<em>abK</em>1−1pi\text{p}<em>i' \propto K</em>1 H<em>{ab} K</em>1^{-1} \text{p}_i

    • Used for image rectification, registration, and camera motion estimation.

Image Preliminaries

  • Image as a 2D grid of grayscale pixels (0-255).

Background: Signal and Image Processing

  • Digital signal processing (DSP): numerical manipulation of signals.

  • Image processing applications: enhancement, filtering, compression.

Computational Hierarchy

  • Levels: Image Capture, Image Processing, Image Analysis, Image Understanding, Object/Scene Models.

  • Example in Autonomous Vehicles: Images → Signals → Preattentive Cues → Lanes, vehicles, pedestrians.

Definitions

  • Pattern Recognition: Measurement Vector → Classified Vectors (not necessarily an image).

  • Computer Vision: Image Derived Measurements + Models → Recognition of Objects and Events.

Computer Vision vs. Graphics

  • Computer Vision: 3D Scene → 2D Image (ill-posed inverse problem).

  • Graphics: 2D Image → Pixel Matrix.

Image Processing Examples

  • Histogram Equalization and Intensity Transforms.

  • Hough Transform.

Machine Learning & Computer Vision

  • Deep Learning Approaches significantly improved ImageNet performance.

Next Lecture

  • Image Acquisition and Representations.