Abstract: This paper studies symmetry detection in natural images, which is a challenging problem in computer vision. To differentiate symmetry and non-symmetry in natural images is intractable due to the large variation in objects and the cluttered scene. To address this problem, two types of edge features motivated by the fact that symmetries are complementary to edges are proposed. These two types of features are both defined on pairs of edges to search pairs of edges with consistent high strength and equal distances to symmetries, respectively. The proposed edge features at multiple scales and orientations
and integrate them with low level cues (color, brightness and texture) under a multiple instance learning framework to detect symmetries are extracted. The experimental results on SYMMAX300 dataset demonstrate that both proposed edge features can improve performance of symmetry detection.

Key words: edge, low level feature, multiple instance learning, symmetry axis, symmetry detection