Mid-Level Vision and Object Recognition Notes
Mid-Level Vision and Object Recognition Notes
Course Information: 01:830:301, Spring 2025, Rutgers University
Page 1: Introduction
Course focused on object recognition and vision complexity.
Page 2: Understanding Object Recognition
Key Questions: How do we recognize different images of the same object? For example, how do we see that all four images are houses?
Page 3: The Challenges of Object Recognition
Current understanding of human visual systems’ ability to recognize objects is limited.
Difficult to create receptive fields (RFs) related to complex items like houses, as they need to accommodate for diversity in perspective, size, and color.
Page 4: Object Recognition in the Brain
Objects are recognized through a series of neural transformations:
Retinal Ganglion Cells: Detect spots.
LGN (Lateral Geniculate Nucleus): Works with these spots.
Primary Visual Cortex (V1): Processes bars.
Extrastriate Cortex: Areas involved in further visual processing (V2, V3, V4, Inferotemporal Cortex).
Page 5: Visual Pathways
Two main processing streams:
Dorsal Pathway: “Where” – involved in spatial awareness.
Ventral Pathway: “What” – involved in object recognition.
Page 6: Cortical Processing Beyond V1
There is a poor understanding of how cortical processing operates post-V1.
Observations of receptive field sizes adjusting and showing position invariance.
Page 7: Boundary Ownership
Discussed sensitivity to boundary ownership particularly in V2, showing how the visual system identifies edges.
Tuning for complex shapes occurs in area V4.
Page 8: Specialized Areas of the Cortex
Different areas of the brain are specialized for recognizing specific features:
PPA: Parahippocampal Place Area - recognizes places.
FFA: Fusiform Face Area - focuses on faces.
EBA: Extrastriate Body Area - identifies body parts.
Page 9: Mid-Level Vision
Refers to processing that occurs between feature detection and object recognition.
Examines how edges, shapes, and surfaces are visually interpreted.
Page 10: Edge Detection
Computer vs. Human Performance: Computers often struggle with accurately detecting edges in images.
Discusses illusory contours – perceptions of edges without actual physical changes present.
Page 11: Two Early Schools of Thought
Structuralism: Asserts perception is formed from basic sensations.
Gestalt: Argues the whole is greater than the sum of its parts.
Page 12: Gestalt Grouping Rules
Key rules include:
Good Continuation
Similarity
Proximity
Common Region
Common Fate
Synchrony
Page 13: Figure and Ground Perception
Understanding which parts of an image represent foreground versus background is key to object recognition. Examples of guiding principles include:
Surroundedness
Size
Symmetry
Parallelism
Convexity
Page 14: Ambiguity in Perception
Describes how perception resolves conflicts utilizing perceptual “committees”. These committees integrate differing opinions and rules to arrive at a consensus.
Page 15: Accidental Viewpoints
The distinction between accidental (unusual) and non-accidental (typical) viewpoints informs how we interpret visual data.
Page 16: Templates vs. Cognitive Models
Discusses the challenge of creating stable templates for object recognition over varying perspectives.
Page 17: Viewpoint Invariance
Some models propose that certain features remain invariant despite the observer’s positional viewing changes.
Page 18: Adversarial Examples in Machine Vision
Highlights vulnerabilities in AI and machine vision systems to manipulated inputs that mislead recognition algorithms.
Page 19: Applications and Example Outputs
Demonstrates how advanced neural networks translate visual inputs into contextually relevant descriptions.
These notes are structured to give a clear overview of the concepts presented in the material on object recognition while covering the complexities associated with human vision and machine learning approaches to recognition.