Abstract
In this report we describe "InnerSpec", an approach for symmetric object detection that is based both on the computation of a symmetry measure for each pixel and on gradient information analysis. The symmetry value is obtained as the energy balance of the even-odd decomposition of an oriented square patch with respect to its central axis. Such an operation is akin to the computation of a row-wise convolution in the midpoint. The candidate symmetry axes are then identified through the localization of peaks along the direction perpendicular to each considered angle. These axes are finally evaluated by computing the image gradient in their neighborhood, in particular checking whether the gradient information displays specular characteristics.
Introduction
Symmetry is a fundamental cue in computer vision, essential for tasks ranging from object recognition to scene understanding. However, detecting reflection symmetry in "natural" images—those containing complex backgrounds, uneven lighting, and occlusions—remains a significant challenge. Traditional mathematical models often fail due to two primary issues: perceptual mismatch, where lighting or noise breaks mathematical symmetry despite human perception of it, and meaningless symmetry, where background textures (like ripples or brickwork) produce high symmetry scores that are irrelevant to the primary object.
InnerSpec addresses these hurdles by moving beyond simple pixel-matching, incorporating a perceptual approach that leverages structural gradient information to distinguish meaningful objects from background noise.
Methodology
The InnerSpec pipeline is divided into three distinct stages designed to transform raw image data into a ranked list of symmetry axes.
1. Symmetry Computation (The "Symmetry Stack")
The algorithm rotates the input image through 90 different angles in 2-degree increments. For each orientation, it calculates a symmetry measure for every pixel using a row-wise auto-convolution of local patches. This process evaluates how well a patch mirrors itself across a central axis, resulting in a 3D "symmetry stack" that represents potential symmetry at every location and angle.
2. Symmetry Axes Identification
Rather than simply selecting the highest global scores, the system identifies local "peaks" within the symmetry maps. A connectivity analysis, similar to a flood-fill algorithm, links these peaks into candidate symmetry segments. These segments are then projected back into the original image's coordinate system for final evaluation.
3. Gradient Refinement
To filter out false positives, InnerSpec employs a secondary check based on image gradients (edges). This "secret sauce" analyzes two specific criteria:
- Specular Magnitude: Ensuring edges on opposite sides of the axis have similar strengths.
- Specular Direction: Verifying that edge directions mirror each other (e.g., a top-right edge on the left must correspond to a top-left edge on the right).
A final score combines the raw symmetry measure, segment length, and gradient specularity to rank the most probable axes.
Experimental Evaluation
InnerSpec was evaluated using a subset of the training data from the 2013 CVPR Symmetry Detection Competition, a standard benchmark featuring natural images with complex backgrounds.
The algorithm demonstrated high stability across various window sizes ($n_p$). As shown in the Precision-Recall (PR) curves, a window size of $n_p = 60$ provided the optimal balance of precision and recall. The integration of gradient processing was found to dramatically aid the system in correctly placing symmetry axes on foreground objects while ignoring background patterns.
Conclusion
InnerSpec demonstrates that "raw data symmetries" are insufficient for robust detection in natural scenes. By incorporating gradient information, the algorithm mimics human perception, focusing on structural boundaries rather than just pixel values. While the computational cost of generating a 90-slice symmetry stack is high, the resulting robustness against noise and background clutter marks a significant step forward in automated symmetry detection. Future work will focus on optimizing the computational complexity and refining gradient weighting for even smaller or more cluttered objects.
Summary updated