Local Feature Detectors MCQ
Harris corners, SIFT, and ORB—scale and rotation aware keypoints for matching and tracking.
Harris Corner Detector MCQ
Harris corner detector
Harris analyzes local image derivatives summed in a window: both eigenvalues large → corner; one large → edge; both small → flat region.
Response trick
The Harris response avoids explicit eigen decomposition using determinant and trace of M.
Interpretation
det & trace
det(M)=λ1λ2, trace(M)=λ1+λ2—combined into R with empirical k ≈ 0.04–0.06.
Shi-Tomasi
Uses min(λ1,λ2) as corner score—often more stable for tracking.
Window size
Larger windows smooth noise but reduce localization; Gaussian weighting is common.
Not scale-invariant
Classic Harris is not intrinsically scale invariant—multi-scale or blob detectors address that.
Eigenvalue diagram
Corner: both λ high · Edge: one λ high · Flat: both λ low
SIFT (Scale-Invariant Feature Transform) MCQ
SIFT overview
SIFT finds scale-space extrema of Difference-of-Gaussian, refines location/scale, assigns dominant orientation(s), then samples gradient histograms into a 128-dimensional vector.
Why DoG?
DoG approximates scale-normalized LoG extrema—efficient multi-scale blob/corner detection.
Stages
Extrema
Compare pixel to 26 neighbors in scale-space cube; reject low contrast / edge-like points.
Orientation
Histogram of gradient orientations weighted by magnitude; peaks create multiple oriented features.
Descriptor
4×4 spatial grid × 8 orientation bins, normalized for illumination robustness.
Patents / use
Historically patent encumbered; now widely usable in research and OpenCV—know licensing for products.
Descriptor layout
16 cells × 8 directions → 128 values (with truncation/normalization steps)
ORB (Oriented FAST and Rotated BRIEF) MCQ
ORB features
ORB combines FAST corners, a simple orientation from intensity moments, and rotation-aware BRIEF bit patterns—optimized for speed on CPUs and embedded devices.
Why binary?
Popcount Hamming distance is extremely fast; storage is compact—tradeoff vs SIFT discrimination.
Pieces
FAST
Compare circle of pixels to center with contiguous arc test—very fast corner detector.
Steered BRIEF
Rotate sampling pattern using θ so bits are rotation invariant.
Learning pairs
ORB selects BRIEF point pairs with lower correlation for better discrimination.
vs SIFT
ORB is faster and free; SIFT often stronger under hard illumination/viewpoint.
Matching cost
Hamming distance between 256-bit strings