UCAS-AI Academy-Computer Vision Special Course- Lecture 1-Course Notes

Course Introduction

• Online lecture + Q & A
• Programming assignment + literature reading assignment 30%
• Class opening 70%

What is computer vision

• Vision: To know what is where by looking
• Perceived uses:
  • Adapt to the environment
  • Control action
• Computer Vision: a discipline that studies visual perception
• Perception: Analysis of feeling information
• Cognition: the process of acquiring knowledge
• The core scientific problem of perception: expression and interpretation (not understanding)
• Computer vision Mo table: build a computer vision system with versatility and flexibility like human vision system
• Computer Vision: From Image to Three-bit Scene Expression
• Computer graphics: from 3D scene expression to image
• Visual knowledge expression: image, video, voice-the relationship between visual concept and concept-reasoning

Four important courses of computer vision development

• Marl Computational Vision Theory
  • Computational vision theory: layer-by-layer processing of image information
  • Three levels
    • Computational theory level
    • Expression and algorithm level
    • Algorithm implementation level
  • The main goal of visual perception: constructing the three-dimensional shape expression of the object layer by layer from the image (3D reconstruction)
    • Computational theory-3D geometric description
    • Expression level-three levels of expression (image-primitive-2.5D (observer coordinate system)-3D (object coordinate system expression)
      • Primitive expression-calculating visible surface information-integrating surface depth, orientation, contour and other information-object coordinate system shape expression
    • Algorithm level-edge extraction, stereo matching
    • Implementation level-neural computing or computer
  • The mainstream view of biological vision believes that depth information is unnecessary
  • Human vision includes object vision and spatial vision, the latter requires three-dimensional shape information
  • The basis of reasoning in concept, 3D shape information is also part of the concept
• Active Visual Debate
  • Questioning and criticizing Marl's visual theory-bottom-to-top theory, lack of high-level knowledge feedback guidance, lack of wooden nails and initiative
  • Purpose and initiative can be integrated into Marr's computational vision framework
  • Difficulties in active vision: gaze and feedback
• Hierarchical three-dimensional reconstruction theory
  • Hierarchical reconstruction: image-projective reconstruction (maintain straight line)-affine reconstruction (maintain parallel)-Euclidean reconstruction (maintain vertical)
  • Advantages: fewer optimization variables involved in each step, and high reconstruction robustness
• Learning-based vision
  • Subspace method (manifold)
    • High-dimensional data can be clustered in low-dimensional space
  • Deep learning methods
    • DNN: Strong expressiveness of stacked structure, driven by field data
    • Object recognition-scene understanding (image-video)
    • Feedforward network-feedback network / recurrent network
    • Deep network interpretability
• Marr's three-dimensional concept theory: to recognize objects, the brain must have an expression of the objects, that is, the three-dimensional shape
• Baggio's two-dimensional image model: the brain's expression of objects is a set of two-dimensional image features in different poses
  • Hmax model
• Decalo's hierarchical de-entanglement theory: hierarchical processing, gradually removing interference information irrelevant to the category of objects, to achieve linearly spaced object expression (popular learning ideas)
  • Untangling model
• Conjecture: Inverse generative model of object recognition
  • Restore the parameters of the generated image layer by layer from the graphics (pose, lighting, geometry, texture ...)
  • Image-inverse transform model-image generation parameters-image generation model-image