Computer Vision

a.k.a. CV

Lecture on Computer Vision (Master of Information Science)

• Definition of “CV,” “CG,” and “Image Processing”

  • CV is the process of extracting “shape/appearance/motion/meaning” from images/videos.
  • The opposite of CV is Computer Graphics (CG), which creates “shape/appearance/motion/meaning” from images/videos.
  • Image Processing involves transforming images/videos into different images/videos, such as changing colors.

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

  • The image posted on the Image Processing page presents a slightly different definition.

• The difficulty of CV is that it is not easily understood by the general public.

  • It is difficult to convey the difficulty of understanding that a pet bottle in front of you is indeed a pet bottle.
  • (blu3mo) There is also a perspective of Programming Education that conveys this sense of understanding.
    • What computers can/cannot do.

• Topics

  • Feature point detection and matching
  • Shape reconstruction from motion (generating point cloud data from multiple images)
  • Computational Photography (photography)
  • 3D reconstruction
  • Image recognition (YOLO, etc.)
  • Recently, Fairness (addressing issues where AI exhibits racial bias) is also being discussed.

• Image recognition

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

  • Object detection, semantic segmentation, etc.

  • History

    • SIFT: Local features using the vector of histograms of brightness gradients
    • Bag of Visual Words: Applying the technique of Bag of Words in Natural Language Processing
    • Image datasets: Caltech-101

  • Methods (basic ones)

    • Conversion to one-hot vectors

    • Improvements can be made by using images (kernels) that are suitable for recognition instead of partial image data.

    • Instead of one-hot vectors, more complex representations can be used.

      • It feels similar to the lineage of NLP.

    • The idea is to compress the width and height dimensions of the image while expanding the depth dimension.

      • Hierarchical structure
      • (blu3mo) The same idea as the diagram in the section on CNN.

• (blu3mo) As long as we use datasets collected by humans, image recognition will adapt and overfit to human cognition.

  • Well, that’s the purpose.
  • It becomes a philosophical discussion about whether we can recognize the “thing itself” through image recognition.
    • It is obvious that it cannot be recognized without human labeling.
    • What if we deliberately ignore the existence of humans and try to achieve object detection (or something similar)?
      • Is that Unsupervised Learning, or can we achieve object detection without supervision?
        • We cannot label using natural language.
        • Can we create an intelligent way of perceiving visual information that is not limited to human vision?
          • Can we reach a level that can be called intelligent?
          • In other words, can we call it intelligent if humans cannot understand it?

#computer Science