Computer Vision and Big Data Analysis

中文版：计算机视觉与大数据分析

This lecture connects big data analytics with computer vision tasks: segmentation, detection, classification, and their real-world applications.

Overview

Data contains various descriptive characteristics or patterns known as features. By leveraging huge amounts of data, we can extract reliable patterns to perform computer vision tasks.

Computer Vision Levels

Computer vision can be categorized into three levels that parallel human visual processing:

Low-Level Vision

• Physiology: Retinal ganglion cells → Lateral Geniculate Nucleus → Primary Visual Cortex
  • Simple cells: orientational sensitivity
  • Complex cells: directional sensitivity
• Net effect: Low-level human vision can be partially modeled as a set of multiresolution, oriented filters
• Implementation: Convolution-based filters and filter banks
  • Detection of edges, corners, and other local features
  • Multiple orientations and scales (“filter pyramids”)
• Applications:
  • First stage of segmentation
  • Texture recognition / classification
  • Texture synthesis

Mid-Level Vision

• Physiology unclear; observations by Gestalt psychologists:
  • Proximity, Similarity, Common fate, Common region, Parallelism, Closure, Symmetry, Continuity, Familiar configuration
• Techniques:
  • Clustering based on similarity
• Applications:
  • Segmentation / grouping
  • Tracking
  • Active contours (Snakes) for contour evolution

High-Level Vision

• Human mechanisms and computational mechanisms:
  • Bayesian networks
  • Templates
  • Linear subspace methods (e.g., PCA)
  • Kinematic models
• Motion analysis:
  • Optical Flow / Feature tracking: no constraints
  • Layered Motion: rigid constraints
  • Articulated: kinematic chain constraints
  • Nonrigid: implicit / learned constraints

Computer Vision Tasks

1. Object Detection and Recognition

• Smile detection: Detecting facial expressions
• Object recognition in supermarkets: LaneHawk by EvolutionRobotics — a smart camera flush-mounted in checkout lanes that continuously watches for items under the basket
• Smart cars (Mobileye): Vision systems in high-end BMW, GM, Volvo models; estimated 70% of global car manufacturers adopt the system
• Vision in space (JPL):
  • Panorama stitching
  • 3D terrain modeling
  • Obstacle detection and position tracking

2. Classification and Prediction

Classification is the problem of discriminating between different classes of objects.

General approach:

1. Find examples with known class (training set)
2. Find features that discriminate within and outside the class
3. Create a function that decides the class given the features
4. Apply the function to new examples

Training and Evaluation:

• Training set consists of records with known class labels
• Training set is used to build a classification model
• A labeled test set evaluates the model quality
• The model is then applied to new records with unknown labels

Common Classification Methods:

Method	Description
Decision Tree	Tree-based splitting on attributes
Rule-based Methods	IF-THEN rules
Memory-based Reasoning	Instance-based learning
Neural Networks	Layered computational models
Naïve Bayes / Bayesian Belief Networks	Probabilistic classifiers
Support Vector Machines	Margin-maximizing classifiers

3. Segmentation and Clustering

Segmentation is a sub-domain of computer vision that aims at grouping similar regions or segments of an image under their respective class labels.

• The image is represented as pixels, making segmentation equivalent to grouping pixels
• Image segmentation is an extension of image classification where localization is also performed
• Segmentation is a superset of image classification: the model pinpoints where an object is present by outlining its boundary

Challenges:

• Preliminary methods such as watershed are prone to over- or under-segmentation
• Low-quality segmentation from traditional approaches

Modern approach: Segmentation by using deep learning on thousands of RGB images.

---

Summary

Task	Purpose	Methods
Detection	Locate and identify objects	Feature extraction, template matching, deep learning
Classification	Assign objects to categories	Decision trees, neural networks, SVM, Bayes
Segmentation	Partition image into meaningful regions	Clustering, active contours, deep learning

Related Concepts

• Computer Vision — Low, mid, and high-level processing
• Image Segmentation — Pixel grouping and boundary detection
• Classification — Supervised learning for object categories
• Neural Networks — Deep learning for vision tasks

Sources

• Lecture-5-Computer-vision-and-big-data-analysis