Overview
The HOG descriptor computes image gradients, bins their orientations inside small cells, and normalizes groups of cells into blocks. The result captures edge and silhouette structure for detection tasks.
CoreSource checked
HOG Feature Descriptor
Describes local image shape by accumulating gradient-orientation histograms over cells.
Quick Facts
Compact metadata for implementation and discovery.
- Category
- Image Processing
- Difficulty
- Intermediate
- Family
- Feature Extraction
- Tier
- Core
- Status
- Verified
- Review
- Source Checked
- Sources
- 2
- Formulas
- 0
- Relationships
- 5
- Code example
- Ready
- Visualization
- Ready
- Aliases
- Histogram of Oriented Gradients
Core Concepts
image feature descriptionfeature extractiongeometric predicatefeature matchingimage processing
Quick facts
Complexity
Runtime and memory behavior at a glance.
Best
O(n)
Average
O(n log n)
Worst
O(n^2)
Space
O(n)
When to Use
- Use when edge and silhouette structure are more important than texture or color.
- Use when a classical descriptor is preferred over learned features.
When Not to Use
- Avoid when viewpoint, scale, or lighting varies beyond what the descriptor can tolerate.
- Avoid when end-to-end learned features are available and substantially more accurate.
What It Does
- Computes local image gradients
- Accumulates orientation histograms in cells
- Normalizes neighboring cells into descriptor blocks
Use Cases
- Object and pedestrian detection
- Classical feature extraction pipelines
- Comparing local shape patterns across image windows
Inputs
- Grayscale or color image grid I(x, y)
- Cell size, block normalization stride, and orientation bin count
Outputs
- Normalized histogram of oriented gradient (HOG) feature vector
- Spatial cell gradient magnitude/orientation maps
Pseudocode
- 1.Compute horizontal and vertical image gradients
- 2.Convert gradients to magnitude and orientation
- 3.For each cell, accumulate weighted orientation votes
- 4.Group neighboring cells into blocks
- 5.Normalize each block and concatenate descriptor values
Canonical Example
- Input
- a 16x16 grayscale patch containing a vertical edge
- Output
- histogram bins with strong vertical-edge orientation response
- Explanation
- Gradients near the edge vote into orientation bins, and block normalization scales the local descriptor.
Implementation Notes
- Define gradient filters, orientation bins, cell size, block size, and normalization method explicitly.
- Use tiny synthetic edge patches to check binning and interpolation.
- Keep descriptor layout stable so trained classifiers receive consistent features.
Common Pitfalls
- Using inconsistent cell, block, or stride sizes makes descriptors incomparable.
- Skipping normalization makes the descriptor too sensitive to contrast.
- Testing only centered objects hides detection-window alignment errors.
Interactive Mechanism Walkthrough
Spatial Primitive & Sweep-Line CanvasShow accumulating local gradient orientation histograms over cells on a small image feature description example.
Step 1 of 5: Compute horizontal and vertical image gradients
image-processing-mechanismNotes
HOG describes local gradient structure; it is usually paired with a classifier or detector window.
Sources (2)
What each reference supports for this entry.
Histograms of Oriented Gradients
PaperOriginal HOG reference for gradient orientation histograms and block normalization.
Open sourceComputer Vision: Algorithms and Applications
BookComputer-vision textbook support for HOG Feature Descriptor's feature, matching, flow, or multiscale image-analysis workflow.
Open sourceExplore Next
Good next jumps from the relationship graph, ranked for useful discovery.
Alternative toORB
Image Processing · Core
ORB is a binary keypoint descriptor, while HOG is a histogram-based window descriptor.
Compare with thisOften paired withTemplate Matching
Image Processing · Core
HOG features and template matching are both used in classical detection pipelines.
Compare with thisRelationships
View full relationship pageConceptual neighborSIFT
Both describe local gradient orientations, but SIFT builds keypoint descriptors while HOG often describes dense windows.
Often paired withTemplate Matching
HOG features and template matching are both used in classical detection pipelines.
Alternative toORB
ORB is a binary keypoint descriptor, while HOG is a histogram-based window descriptor.
Related Algorithms
Tags
Image Processing