CDS 6334 - Visual Image Processing

Lecture 9: Local Invariant Features

1. Why Local Invariant Features?

Local invariant features describe small image regions that remain recognizable under transformations.
Local features are more robust than global image representations.
🧠 Match small regions instead of entire images.

2. Motivation

Image matching becomes difficult when images differ in scale, rotation, viewpoint or lighting.
Local features improve robustness against these changes.

3. General Matching Pipeline

  1. Detect keypoints
  2. Define local regions
  3. Extract descriptors
  4. Match descriptors
🧠 Detect → Describe → Match

4. Main Components of Local Features

Stage Purpose
Detection Find interest points
Description Create feature vectors
Matching Find correspondences

5. Requirement: Repeatability

The same keypoints should be detected independently in different images.
Repeatable detectors are necessary for successful matching.

6. Requirement: Distinctiveness

Each feature should have a unique description.
Distinctive descriptors reduce false matches.

7. Desired Properties

Exam Keyword:
Invariance

8. Invariance

Features should remain stable under image transformations.
Types:
  • Translation
  • Rotation
  • Scale
  • Illumination

9. Interest Point Detection

Detection identifies important image locations suitable for matching.
Corners are among the most useful interest points.

10. Why Corners?

Corners contain strong intensity changes in multiple directions.
Corners are repeatable and distinctive.

11. Good Corner Characteristics

12. Harris Corner Detector

Harris detects corners using image gradient information.
Uses a 2×2 matrix of image derivatives.
Important Method:
Harris Corner Detector

13. Harris Corner Principle

A corner exists when intensity changes significantly in two directions.
🧠 Large changes in X and Y → Corner

14. Eigenvalues Interpretation

Eigenvalues Meaning
Both Small Flat Region
One Large Edge
Both Large Corner

15. Harris Detection Steps

  1. Compute corner response
  2. Apply threshold
  3. Perform non-maximum suppression
🧠 Response → Threshold → Local Maxima

16. Harris Properties

Property Supported?
Translation Invariant Yes
Rotation Invariant Yes
Scale Invariant No
Exam Favourite:
Harris is NOT scale invariant.

17. Need for Scale Invariance

The same object may appear at different sizes in different images.
Interest points must be detectable across scales.

18. Scale-Space Concept

Features are searched across multiple image scales.
🧠 Search in position AND scale.

19. Blob Detection

Blob detectors locate regions with strong contrast differences.
Method:
Laplacian of Gaussian (LoG)

20. Characteristic Scale

The best scale is the one producing the maximum Laplacian response.
Interest points are local maxima in both position and scale.

21. Difference of Gaussians (DoG)

DoG approximates the Laplacian of Gaussian efficiently.
🧠 DoG ≈ LoG but faster.

22. Feature Description

A descriptor summarizes the appearance around a keypoint.
Descriptors are used for matching.

23. Raw Patch Descriptors

A simple descriptor uses pixel intensities directly.
Highly sensitive to rotation and translation.

24. SIFT

Scale Invariant Feature Transform (SIFT) is a robust local descriptor.
Important Method:
SIFT

25. SIFT Procedure

  1. Detect keypoints using DoG
  2. Find dominant orientation
  3. Rotate patch to canonical orientation
  4. Normalize scale
  5. Create descriptor
🧠 Detect → Orient → Normalize → Describe

26. Dominant Orientation

SIFT computes a histogram of gradient directions.
Orientation with highest vote becomes the dominant direction.

27. SIFT Descriptor Structure

4 × 4 histogram grid
8 orientation bins per grid
Descriptor Length:
4 × 4 × 8 = 128 values
Exam Favourite:
SIFT Descriptor = 128 Dimensions

28. Properties of SIFT

🧠 One of the most successful feature descriptors.

29. Applications of Local Features

30. Final Exam Summary

Most Important Points

  • Local Features: Detect, describe and match image regions.
  • Requirements: Repeatability and distinctiveness.
  • Invariance: Translation, rotation, scale and illumination.
  • Corners: Strong intensity changes in multiple directions.
  • Harris: Corner detector using image gradients.
  • Harris Properties: Translation & rotation invariant but NOT scale invariant.
  • Scale-Space: Detect features across multiple scales.
  • LoG: Blob detector.
  • DoG: Efficient approximation of LoG.
  • SIFT: Scale Invariant Feature Transform.
  • Dominant Orientation: Provides rotation invariance.
  • SIFT Descriptor: 4×4×8 = 128 dimensions.
  • Applications: Recognition, tracking, mosaicing, 3D reconstruction.