Histograms & Feature Intro — Interview Q&A

Answer: Remaps intensities so output histogram is more uniform—spreads contrast using the cumulative distribution function (CDF) as a transform.

Answer: Count of pixels at each intensity level (per channel). For 8-bit gray, 256 bins—shows under/over exposure and bimodality.

Answer: Transform T(k) maps input level k using normalized CDF × (L−1)—monotonic mapping preserves order, spreads occupied intensity ranges.

Answer: Single mapping from whole-image histogram—fast but can fail with spatially varying illumination (washes out regions).

Answer: Uniform use of gray levels maximizes entropy in a discrete sense—improves perceptual contrast when information was compressed into narrow intensity band.

Answer: Stretching flat regions spreads quantization noise across more levels; CLAHE limits contrast locally to reduce this.

Answer: Contrast Limited Adaptive HE: run HE on small tiles, clip histogram before equalization to limit slope, interpolate tile borders—handles uneven lighting better than global HE.

Answer: Caps histogram bin heights before redistribution—limits maximum local contrast boost; higher clip → stronger enhancement but more noise.

Answer: Small tiles: local adaptation but blocking artifacts if interpolation weak; large tiles: approaches global HE.

Answer: Independent per-channel HE changes hue/saturation—unnatural colors. Prefer luminance-only or LAB L channel.

Answer: Convert to LAB, equalize L* only, convert back—preserves chroma better than RGB HE.

Answer: Linearly maps [min,max] to full range—simpler, no CDF; HE is nonlinear full remap based on distribution shape.

Answer: Gamma is parametric power curve; HE is data-driven. Gamma doesn’t require histogram computation; HE adapts to image statistics.

Answer: Match histogram to a target distribution via mapping through CDFs—generalization of equalization (uniform target).

Answer: Marks pixels whose colors match a model histogram—used in classic CamShift / skin detection pipelines.

Answer: Improve tissue visibility; must avoid misleading diagnosis—sometimes CLAHE on X-ray/CT views; DL often learns normalization end-to-end now.

Answer: Less common if dataset is large; can help low-light inputs or classical pre-steps; batch norm and augmentation reduce reliance.

Answer: Mapped values rounded to 256 levels—true uniform continuous histogram impossible; some bins may stay empty.

Answer: cv2.equalizeHist for grayscale; cv2.createCLAHE(clipLimit, tileGridSize) for CLAHE.

clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
y = clahe.apply(gray)

Answer: When preserving absolute photometry matters, or scene already high contrast—HE can clip highlights or crush semantic color cues.

Answer: A salient image patch with a keypoint (location, scale, orientation) and often a descriptor vector summarizing local appearance for matching.

Answer: Detector finds stable interest points; descriptor encodes local neighborhood for similarity—can mix (e.g. Harris corners + SIFT descriptor in some pipelines).

Answer: High gradient in multiple directions—well localized, repeatable under small viewpoint/light changes vs flat regions or straight edges.

Answer: Extremum in scale-space (LoG/DoG)—captures roundish regions; complementary to corners for texture-poor scenes.

Answer: Detect+describe at multiple scales or with scale-normalized patch so matching works across zoom—SIFT pyramid, ORB octave pyramid.

Answer: Assign dominant orientation from gradient histogram and rotate patch canonical frame—or use rotation-invariant descriptors (some trade distinctiveness).

Answer: Regions that deform predictably under affine viewing of planar surfaces—MSER, Harris-Affine family; stronger than similarity for wide baselines.

Answer: Normalized cross-correlation over patches—brightness/contrast normalized SSD-like score; expensive vs sparse keypoints.

Answer: Reject match if distance to nearest neighbor is not sufficiently smaller than second-nearest—reduces ambiguous matches.

Answer: Accept match only if a is nearest to b and b nearest to a—simple filter for symmetric uniqueness.

Answer: Estimates geometric model (F/E/H) while rejecting outliers from incorrect matches—essential for robust pose and stitching.

Answer: Quantize descriptors to vocabulary clusters; image → histogram of words—classic image retrieval / classification before deep CNNs.

Answer: Track keypoints frame-to-frame with KLT, optical flow, or re-detect+match—balance drift vs redetection.

Answer: Sparse landmarks for bundle adjustment; need repeatable detection and robust data association across frames.

Answer: Same real-world point detected under noise, blur, and viewpoint change—measured by overlap of keypoint regions on benchmark sequences.

Answer: Descriptor separates correct matches from distractors—low false match rate at fixed threshold.

Answer: Keypoints disappear; need robust matching, wide baseline tolerance, or dense methods / learning-based segmentation.

Answer: For binary descriptors (BRIEF, ORB)—XOR + bit count; fast with POPCNT hardware.

Answer: Fast Library for Approximate Nearest Neighbors—speeds k-NN on high-dim descriptors with trees or LSH; trade accuracy for speed.

Answer: SuperPoint, LIFT, etc.—CNNs predict keypoints+descriptors end-to-end; outperform classical on some benchmarks with enough data.