Symmetry is a common geometric invariant that exists in nature and man-made environments. In spite of its ubiquitous nature, the (1) detection of symmetry and the (2) extraction of symmetrical structures that supports the symmetry remains a challenge in Computer Vision. In this work, we present a complete approach that addresses these two processes together, specifically for the case of curved reflection symmetries found in articulated structures. For the detection of curved reflection symmetries, we train a Structured Random Forest (SRF) classifier that associates local symmetrical measures computed via fast histogram comparisons  from color, texture, Gabor and spectral features with curved symmetries. Next, we embed the detected curved symmetries into a 5-way Markov Random Field (MRF) representation of the image edges so that symmetry is enforced in the final segmentation using graph-cuts. As symmetry is enforced locally, our approach is able to handle 1) multiple branches and 2) approximate symmetries which are common in real images.
Questions? Please contact cteo "at" umd dot edu