Computer Vision

Also known as Computer Vision (CV)

Lecture on Information Science

• Definitions of “CV,” “Computer Graphics (CG),” and “Image Processing”

  • Extracting “shape/appearance/movement/meaning” from images/videos is what CV is about
  • Antonym of CV: Computer Graphics (CG)
    • Creating “images/videos” from “shape/appearance/movement/meaning” is CG
    • Opposite direction of CV
  • Image Processing involves transforming “images/videos” into different “images/videos”
    • Such as changing colors

• (blu3mo) This above seems slightly different from the definition in the book “Digital Image Processing”

  • The image posted on the Image Processing page seems to present a slightly different definition

• The difficulty of CV lies in being hard for the general public to understand

  • It’s difficult to convey the challenge of recognizing a plastic bottle as a plastic bottle right in front of you
  • (blu3mo) There also seems to be a perspective of conveying this kind of understanding in Programming Education
    • What computers can/cannot do

• Topics

  • Feature point detection and matching
  • Shape reconstruction from motion (generating point cloud data from multiple images)
  • Computational Photography
  • 3D reconstruction
  • Image recognition (such as YOLO)
  • Recently, there is also a focus on Fairness (issues where AI exhibits racial bias)

• Image Recognition

  • Combination: Machine Learning x Computer Vision x Natural Language Processing

  • Object detection, semantic segmentation, etc.

  • History

    • SIFT: Local features using histograms of gradient magnitudes
    • Bag of Visual Words: Applying the Bag of Words method from Natural Language Processing
    • Image datasets: Caltech-101

  • Methods (basic ones)

    • To improve,
      • Use images (kernels) suitable for recognition rather than data from parts of images
      • Use more complex representations than One-hot encoding
        • Resembles the lineage of NLP
    • Conceptually, compress the dimensions of images while expanding the depth dimension
      • Hierarchical structure
      • (blu3mo) Similar to the diagram in the CNN section

• (blu3mo) As long as we work with datasets collected by humans, we adapt and overfit to human cognition

  • Well, that’s the goal
  • It might lead to philosophical discussions about whether image recognition relates to recognizing the “thing itself”
    • It’s naturally impossible to recognize without human labeling
    • What if we deliberately ignore human presence and work towards object detection?
      • Would that be Unsupervised Learning, or can we achieve object detection (or something close) without human supervision?
        • Labeling through natural language is obviously not possible
        • Can we create an intelligent way of visual perception other than human?
          • Can we reach a level that can be called intelligent?
          • Or rather, if humans can’t understand it, can we call it intelligent?

#informationscience