Classical versus Deep Mirror-Symmetry Scoring: A Benchmark of Thirteen Methods

This paper presents the first comprehensive benchmark of mirror-symmetry scoring methods, evaluating 13 distinct approaches—ranging from classical hand-crafted features to modern deep-learning backbones—on a unified, statistically grounded protocol. The author recasts these diverse methods into a single computational template: representing an image, comparing it to its mirror reflection, and aggregating the result into a scalar symmetry score. The benchmark uses a reflection-exact harness across nine datasets, employing a chance-anchored discrimination task where a scorer must correctly rank a true symmetry axis above various perturbed (shifted or rotated) negative axes.

The study reveals that symmetry discrimination is primarily driven by mid-scale oriented features. While deep backbones generally achieve the highest performance, the performance gap between the best-performing frozen deep features and a well-tuned classical HOG descriptor is small and often statistically indistinguishable from other high-performing classical measures. Crucially, the HOG descriptor offers a massive computational advantage, executing roughly 300 times faster than deep-network readouts on standard CPUs. The analysis also suggests that the effectiveness of deep models for this task peaks at low-to-mid network stages, indicating that the spatial hierarchy of early-to-mid layers is more critical for symmetry detection than the high-capacity, abstract representations found in deeper layers.

For many practical applications—such as medical imaging, aesthetic analysis, and developmental biology—the ability to quantify symmetry is essential. This work provides an actionable roadmap for practitioners, demonstrating that expensive deep-learning models are not always necessary for symmetry scoring. By releasing the 'imgsym' toolkit, the author provides a standardized framework that allows researchers to select the most efficient and effective method for their specific computational constraints, effectively challenging the assumption that deep features are inherently superior for all visual discrimination tasks.