pyfhog 0.1.1

Fast FHOG feature extraction for Python

PyFHOG - Fast Felzenszwalb HOG Feature Extraction

A minimal Python package that wraps dlib's optimized FHOG (Felzenszwalb Histogram of Oriented Gradients) implementation for cross-platform distribution.

Validated: Perfect correlation (r = 1.0, RMSE = 0.0) with OpenFace 2.2

Features

• Fast: Full C++ SIMD performance via dlib
• Cross-platform: Distributes as wheels for Mac/Linux/Windows
• Simple API: Single function for FHOG extraction
• Compatible: Produces identical output to OpenFace 2.2 (validated r = 1.0)

Installation

pip install pyfhog

Usage

import pyfhog
import numpy as np
import cv2

# Load image (must be RGB format)
img = cv2.imread('face.jpg')
img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

# Extract FHOG features
features = pyfhog.extract_fhog_features(img_rgb, cell_size=8)

# For 112x112 image (OpenFace standard): features.shape = (4464,)
# Formula: (112/8 - 2) * (112/8 - 2) * 31 = 12 * 12 * 31 = 4464
# Note: dlib excludes border cells, hence the -2

FHOG Details

FHOG (Felzenszwalb HOG) is an enhanced version of standard HOG with:

• 31 features per cell (vs 9 for standard HOG)
  • 18 signed orientation bins
  • 9 unsigned orientation bins
  • 4 texture energy features
• Cell size: Typically 8x8 pixels
• Output: Flattened 1D array of features

Why PyFHOG?

PyFHOG solves the cross-platform distribution problem for FHOG extraction:

Approach	Pros	Cons
Pure Python	No compilation	20-40% slower, accuracy risk
Platform binaries	Fast	Hard to distribute, platform-specific
PyFHOG (wheels)	Fast, Cross-platform, Easy install	-

Technical Details

• Based on: dlib's extract_fhog_features() implementation
• SIMD optimized: Uses SSE/AVX on x86, NEON on ARM
• Wheel size: ~200-500KB (minimal overhead)
• Python binding: pybind11

Building from Source

git clone https://github.com/johnwilsoniv/pyfhog.git
cd pyfhog
pip install pybind11 numpy
python setup.py build_ext --inplace
python -m pytest tests/

License

Boost Software License 1.0 (matches dlib's license)

Credits

• dlib: C++ FHOG implementation by Davis King
• OpenFace: Reference implementation by Tadas Baltrušaitis