1. Why Segmentation?
Segmentation divides an image into meaningful regions or objects.
Pixels belonging to the same object should be grouped together.
🧠Segmentation = Partitioning an image into meaningful parts.
2. Objectives of Segmentation
- Group related image features
- Identify objects or regions
- Reduce image complexity
- Prepare for higher-level analysis
Exam Keyword:
Region Grouping
3. Examples of Segmentation Tasks
| Task |
Description |
| Shot Detection |
Group video frames into scenes |
| Region Segmentation |
Separate image regions |
| Object Segmentation |
Extract object boundaries |
| Figure-Ground Separation |
Separate foreground and background |
4. Top-Down vs Bottom-Up Segmentation
| Approach |
Idea |
| Top-Down |
Pixels belong together because they form the same object |
| Bottom-Up |
Pixels belong together because they look similar |
5. Gestalt Theory
Gestalt Theory explains how humans naturally group visual elements.
Common Gestalt Laws:
- Similarity
- Symmetry
- Closure
- Continuity
- Common Fate
🧠Humans group things that look alike or move together.
6. Segmentation Goal
The primary goal is to separate an image into coherent objects or regions.
Similar pixels should belong to the same segment.
7. Thresholding Revisited
Thresholding separates pixels into groups using intensity values.
Pixels above and below a threshold become different segments.
Important Method:
Otsu's Thresholding
8. Motivation for Clustering
Clustering automatically discovers groups within image data.
🧠Clustering finds representative groups without manually selecting thresholds.
9. Clustering Review
Pixels are assigned to the nearest cluster centre.
Goal: Minimize Sum of Squared Distances (SSD).
10. K-Means Clustering
K-Means partitions data into K clusters by minimizing within-cluster variance.
Exam Keyword:
SSD (Sum of Squared Distances)
11. K-Means Algorithm
- Select K
- Initialize cluster centres
- Assign points to nearest centre
- Update cluster means
- Repeat until convergence
🧠Assign → Update → Repeat
12. Segmentation as Clustering
Segmentation can be performed by clustering pixels in feature space.
Different feature choices produce different segmentations.
13. Intensity-Based Segmentation
Pixels are grouped according to grayscale intensity values.
Feature Space:
1-D Intensity
14. Colour-Based Segmentation
Pixels are grouped using colour similarity.
Feature Space:
RGB (3-D)
15. Intensity + Position Features
Spatial coordinates can be combined with intensity information.
Separates regions that share similar colours but occur in different locations.
16. Texture-Based Segmentation
Pixels are grouped according to texture descriptors.
Feature Space:
Filter Bank Responses
17. Superpixels
Superpixels merge similar neighbouring pixels into larger units.
Significantly reduces the number of elements for processing.
🧠Many pixels → Fewer superpixels
18. Images as Graphs
Images can be represented as graphs where pixels are nodes and similarities are edges.
Node = Pixel
Edge = Similarity
19. Graph Cuts
Segmentation is achieved by removing weak connections between regions.
Low similarity edges are good candidates for cuts.
20. Normalized Cuts
A graph-based segmentation algorithm that partitions images using pixel affinities.
Exam Keyword:
Spectral Clustering
21. Felzenszwalb Segmentation
Efficient graph-based segmentation method that merges similar regions.
Steps:
- Compute edge weights
- Sort weights
- Merge similar regions
- Repeat until complete
🧠Small weight = High similarity
22. SLIC Superpixels
Simple Linear Iterative Clustering (SLIC) generates compact superpixels.
Requires only one major parameter:
Number of Superpixels.
23. Why Are Superpixels Useful?
- Reduce computational cost
- Preserve object boundaries
- Enable region-based classification
- Simplify feature extraction
24. Deep Learning Segmentation Methods
| Method |
Main Purpose |
| FCN |
Semantic Segmentation |
| Mask R-CNN |
Instance Segmentation |
| Vision Transformer (ViT) |
Transformer-Based Segmentation |
| SAM |
Prompt-Based Segmentation |
25. Segment Anything Model (SAM)
SAM is a foundation model capable of segmenting objects from prompts.
Evolution:
- SAM 1 → Promptable segmentation
- SAM 2 → Video segmentation & tracking
- SAM 3 → Concept segmentation using text/image prompts
Important Concept:
Zero-Shot Segmentation
26. Final Exam Summary
Most Important Points
- Segmentation: Divide images into meaningful regions.
- Gestalt Theory: Human-inspired grouping principles.
- Thresholding: Basic segmentation using intensity.
- Clustering: Group pixels based on similarity.
- K-Means: Assign → Update → Repeat.
- Feature Spaces: Intensity, colour, position, texture.
- Superpixels: Merge pixels into larger meaningful units.
- Graph Cuts: Remove weak connections between regions.
- Normalized Cuts: Spectral graph segmentation.
- Felzenszwalb: Efficient graph-based segmentation.
- SLIC: Popular superpixel algorithm.
- FCN: Semantic segmentation network.
- Mask R-CNN: Instance segmentation.
- ViT Segmenter: Transformer-based segmentation.
- SAM: Promptable foundation segmentation model.