Attention based: modeling human perception of reflectional symmetry in the wild

Symmetry exists everywhere, presenting itself in various forms and scales. Humans take advantage of symmetry for various tasks, but superb symmetry perception through computational models remains elusive. Symmetry is a salient visual cue that attracts human visual attention. The existing deep learning techniques detect symmetry without attention to the region of interest or understanding various symmetry cues. The proposed novel end-to-end deep learning model leverages the attention mechanism to segment and detect reflectional symmetry (SDRS) in real-world images. The incorporation of attention mechanisms obliging the SDRS Network (SDRS_Net) to focus on the region of interest and approximate the human perception of Reflectional symmetry in the wild. We also present a new Dense and Diverse Human Symmetry Perception (DDHSP) dataset based on human perception of reflectional symmetry that mitigates the limitations of the existing datasets. The SDRS_Net is trained on the DDHSP dataset and quantitatively evaluated against state-of-the- art symmetry detection algorithms. The SDRS_Net reported 0.70 and 0.76 F-scores on the DDHSP and publicly available symmetry test datasets. The quantitative results show that the SDRS_Net is superior to the existing state-of-the-art symmetry detectors.