Give the definition of feature.
A feature is a “meaningful” part of the image.
Features have two main components
– Feature detection: finding a “stable” (easily detectable) point
– Feature description: a description of the surrounding area
What is a feature matching? How can be useful?
- Evaluate features in two images
- Find similar features (good matches)
- Object localization
- Stitching image mosaics
- Scene reconstruction
- Place recognition
- Object detection
- Follow patterns in video flows
How Harris corners are computed?
- Consider a patch in a given position (𝑥𝑖, 𝑦𝑖)
- Consider a displacement Δ𝑥, Δ𝑦
- Similarity is measured by means of the autocorrelation, a function of the displacement
Consider Harris corners detector: describe the auto-correlation matrix.
Matrix
Studying the eigenvalues we get information about the type of patch.
Describe SUSAN corner detector.
– Analyzes a circular window around the point
– No derivatives involved
– Edge+corner detector
– Robust to noise
What is a blob?
Feature!
A blob is a region where
– Considered properties are different from surrounding regions
– Properties are (approximately) constant inside the region
What is the Maximally Stable Extremal Regions (MSER)?
Algorithm that uses blobs!
- Apply a series of thresholds (e.g., one for each gray level)
- Compute the connected binary regions
- Compute some statistics for each region
• E.g.: area, convexity, circularity, … - Analyze how persisten each blob is
How can we create different scales of an image?
Using the N-dimensional gaussian kernel and varying t.
Give the formula!
What is the SIFT and what are its strong points?
- Very reliable keypoint detector and descriptor.
- Image content mapped into local feature
coordinates.
Strong points:
– Local – robust to occlusions
– Distinctive – distinguish objects in large databases
– Dense – many features can be found even on small objects
– Efficiency – fast computation
Describe the SIFT algorithm.
- Scale-space extrema (no local) detection: use Difference of Gaussians (DoG)
- Keypoint localization and refinement: search for maxima and minima of the DoG and refine them using the Hessian matrix (2nd order derivative)
- Orientation measurement: compute gradient magnitude and orientation in the keypoint neighborhood (not with the Sobel operators) - takes also the 2nd peak for duplicating the keypoint
- Descriptor calculation: create histograms and the descriptor vector
What are the Haar features for face detection?
Rectangular filters.
Local feature: subtract the sum of pixels in the white areas from the sum of pixels in the black area.
Viola and Jones algortihm.
- Define Haar features
- Weak learners working on Haar features
- Boosting for combining multiple weak learners
- Cascade of classifiers
Other SIFT-based features.
- PCA-SIFT
- SURF (Speeded Up Robust Features)
- GLOH (Gradient Location-Orientation Histogram)
