HOG (Histogram of Oriented Gradients) MCQ
Compute gradients, accumulate signed orientation votes per cell, concatenate block-normalized histograms for a fixed-length template.
Gradients
Gx, Gy
Cells
8×8 etc.
Blocks
Normalize
Pedestrian
Window
HOG descriptors
HOG summarizes local edge orientations in a dense grid. Block-wise contrast normalization makes the descriptor robust to illumination while preserving shape cues—classic for rigid pedestrian templates + linear SVM.
Why blocks?
L2 normalize over overlapping blocks of concatenated cell histograms to cancel local lighting gradients.
Pipeline
Gradients
Finite differences with optional Gaussian pre-smoothing; magnitude weights orientation votes.
Cells
Unsigned orientations binned (e.g., 9 bins over 0–180°) per cell.
Blocks & stride
2×2 cells per block with stride controls overlap and final dimensionality.
Modern use
CNNs largely replaced hand-tuned HOG+SVM, but HOG teaches gradient histogram ideas still used inside networks.
Sliding window
Fixed aspect window scans image; each window → one HOG vector → classifier score