HOG descriptor matches deep learning for symmetry scoring at 300x speed
Classical HOG features rival frozen deep networks on mirror-symmetry, running 300x faster on CPU.
A comprehensive benchmark published on arXiv by Maximilian Woehrer evaluates 13 methods for quantifying how mirror-symmetric an image is about a given axis—a task critical for visual aesthetics, medical imaging, and industrial inspection. The study compares nine existing methods and four novel ones, spanning classical computer vision features (e.g., HOG) to frozen deep neural network embeddings, across four single-axis and five multi-axis datasets under a rigorous, statistically validated protocol.
Surprisingly, the classical Histogram of Oriented Gradients (HOG) descriptor performs nearly as well as the best frozen deep network readouts, trailing by only a small (but significant) margin, and is not statistically separable from the runner-up CNN-filter measure. More importantly, HOG runs ~300x faster on CPU than deep alternatives. The analysis reveals that discrimination power concentrates in mid-scale oriented features—both HOG and early-to-mid stages of deep networks capture this effectively. The paper releases the 'imgsym' open toolkit, suggesting that for symmetry scoring, expensive deep features may be overkill when a well-tuned classical descriptor suffices.
- HOG descriptor trails best frozen deep network by a small margin, with no statistical difference from the runner-up CNN-filter method.
- Classical HOG runs ~300x faster on CPU than deep learning inference, enabling real-time symmetry scoring without GPUs.
- Performance concentrates in mid-scale oriented features; deep networks peak at low- to mid-stages, while HOG peaks at mid cell size.
Why It Matters
Challenge to deploy deep learning for symmetry tasks—classical computer vision can match performance at a fraction of the cost.