Skip to content

02 Vision

1. ROC Curves Recap

  • Two ways to construct an ROC curve:

    1. Use one observed point (hit & false alarm rate) + theoretical model.

    2. Collect multiple data points by varying criterion (payoffs or confidence ratings).

  • Confidence ratings allow multiple cutoff points to generate a curve.

  • Area under curve (A) indicates sensitivity.

2. Early Vision Lecture Roadmap

  • What it means to study a sensory modality.

  • Eye structure & function.

  • Retina structure & function.

  • Receptive fields & lateral inhibition.

  • Pathway to brain (LGN).

  • Visual cortex (V1).

  • Spatial frequency analysis.

3. Fundamentals of Studying Sensory Systems

  • Key questions: receptors, physical information transduced, peripheral processing, functional modules, algorithms.

  • For vision:

    • Receptors: rods & cones.

    • Input: photons trigger molecular reactions.

    • Peripheral processing: inter-neural interactions → circular receptive fields.

    • Functional modules: object recognition (“what”) & localization (“where”), with distinct pathways.

    • Algorithms: many, often produce illusions.

4. Eye & Focus

![[Screenshot 2025-09-04 at 9.43.31 AM.png]] - Goal of the outer surface (cornea角膜 & lens): focus light onto retina (视网膜). - Focal problems = blur (too near/far). - Lens power (diopters) = 1/focal length (m). - Accommodation = lens shape change by ciliary muscles. - Visual angle: measure of retinal size (thumb rule ≈ 2° at arm’s length).![[Screenshot 2025-09-04 at 9.52.23 AM.png]]

5. Retina Structure

![[Screenshot 2025-09-04 at 9.55.26 AM.png]] - Photoreceptors: ~90M rods (scotopic, sensitive, periphery) & ~5M cones (photopic, color, fovea小凹，尤指视网膜的中央凹). - Convergence patterns: - Rods → diffuse bipolar cells → M ganglion cells (sensitive, low acuity). - Cones (fovea) → midget bipolar cells → P ganglion cells (high acuity, detail, color). - Ganglion cells: ~1M, function as spot detectors, divide into magno vs parvo pathways. - Relative acuity: fovea = high detail, periphery = low detail.

6. Receptive Fields & Lateral Inhibition

  • Receptive field = area & pattern of stimulation affecting cell response.
  • Bipolar & ganglion cells have on-center/off-surround or off-center/on-surround receptive fields.
  • Lateral inhibition: excited receptors inhibit neighbors → edge enhancement.
  • Functions: sharpen borders, enhance contrast, produce illusions (Mach bands, Hermann grid, scintillating grid

7. Leaving the retina: Pathways toward the brain

  • Signals leave retina via optic nerve.
  • Cross at optic chiasm → each hemisphere processes contralateral visual field.
  • LGN (lateral geniculate nucleus in thalamus): layered structure, separated by eye of origin and ganglion cell type (M, P).
  • LGN relays to V1 but also receives strong feedback from V1.

8. Visual Cortex (V1)

  • Retinotopic map: V1 spatially represents retina.

  • V1 hypercolumns: contain orientation-selective cells, ocular dominance columns, and blobs (color role unclear).

  • Receptive fields evolve:

    • Retina/LGN: spots.

    • V1: edges, orientation.

    • Simple vs complex cells differ in selectivity (position + orientation vs just orientation).

  • Multiple pathways:

    • V1 (pattern, motion, depth).

    • Superior colliculus (head orienting).

    • Suprachiasmatic nucleus (circadian).

    • Other reflexive visual functions.

9. Spatial Frequency Analysis

  • Vision analyzes images by spatial frequency (edge density).

  • Spatial frequency selective cells appear as early as retinal ganglion cells.

  • Human sensitivity curve: most sensitive to intermediate frequencies.

  • Complex images can be decomposed into sine waves (Fourier analysis).

  • Filtering:

    • High-pass = edges, detail.

    • Low-pass = blur, coarse shapes.

  • Demonstrations:

    • Lincoln/Julesz image (low-pass reveals identity).

    • Hybrid images (different perception at near/far or early/late).

  • Scene perception proceeds coarse-to-fine (low → high frequency).