Bilateral symmetry detection in visual arts

Collaboration with CIEREC, Univ. Saint-Etienne

Bilateral symmetry is a very important intermediate feature in some computer vision applications such as aesthetic analysis, object detection, depth estimation or medical image processing. It corresponds to significant region where pixel colors are symmetrically arranged wrt. a local symmetry axis. In visual arts, bilateral symmetry it connected to the formal principle of balance which describes how the main patterns are arranged symmetrically or asymmetrically to create the impression of equilibrium inside the image.

State-of-the-art methods combine feature point detection, pairwise comparison and voting in Hough-like space. In spite of their good performance, they fail to give reliable results over challenging real-world and artistic images.

In this work, our contributions are:

• Introducing a new local edge descriptor and an associated symmetry measure.
• Formalizing symmetry detection as a kernel density estimation problem.
• Proposing a symmetry dataset based on aesthetic analysis.

The global principle of our method is given below and consists of:

• Wavelet-based feature extraction at different scales and determination of local orientation and texture histogram.
• Pair-wise feature triangulation and symmetry weight computation in the polar coordinate system.
• Voting representation based on weighted pairs via linear-directional kernel density estimation.
• Axis selection by searching for maximum peaks, and spatially defined by the convex hull of the voting features.

Quantitative and qualitative comparisons show a substantial advantage for our proposed method on different types of public datasets. Our algorithm was awarded in ICCV’17 workshop challenge Detecting symmetry in the wild at the2^nd position for the challenge 2D symmetry multiple and 3^rd position for the challenge 2D symmetry single (see results here ).

For the purpose of visual art annotation, we labeled global-axis symmetry ground-truth for 253 images extracted form Aesthetic Visual Analysis (AVA) dataset. The corresponding AvaSym dataset is publicly available here.

The figure below presents some qualitative results comparing our approach over three concurrent ones.

Some qualitative results comparing our approach over three concurrent ones.

Publications
[1] M. Elawady, C. Barat, C. Ducottet, P. Colantoni, Global Bilateral Symmetry Detection Using Multiscale Mirror Histograms,  In International Conference on Advanced Concepts for Intelligent Vision Systems, pp. 14-24, 2016

[2] M. Elawady, C. Ducottet, O. Alata, C. Barat, P. Colantoni, Wavelet-based Reflection Symmetry Detection via Textural and Color Histograms , ICCV 2017, IEEE International Conference on Computer Vision Workshop Detecting Symmetry in the Wild, Venice, Italy, 2017

[3] M. Elawady, O. Alata, C. Ducottet, C. Barat, P. Colantoni, Multiple Reflection Symmetry Detection via Linear-Directional Kernel Density Estimation , CAIP 2017, 17th International Conference on Computer Analysis of Images and Patterns, pp. 344-355, Ystad, Sweden, 2017

< Back