Learning-based Symmetry Detection From Real World Images

Symmetry as one of the important global features of objects,has long been a research hotspot in computer vision.Symmetry detection aims to extract symmetric information from given objects.The information usually includes object locations and categories,such as translation,rotation and bilateral symmetries;and the geometric properties of different categories such as lattice for translation symmetry,rotational center and supporting regions for rotation symmetry and reflection axes for bilateral symmetry.Many problems exists in traditional symmetry detection methods: it is hard to detect all three major categories of symmetries under a unified frame work,also it is difficult to keep robust detections due to the complexity of nature images,and most traditional methods are time-consuming,which makes them hard to be applied into practical use.To overcome these disadvantages,in this paper,a learning-based symmetry detection method is proposed based on current accomplishments of deep convolutional neural networks.A few improvements are made in this paper for better symmetry detection and geometric properties extraction.The main work of this paper includes:(1)A symmetry prior is introduced into the proposed network based on the properties of symmetry patterns.For certain convolutional layers,symmetric constrains are added during the process of kernel weights update.The symmetric kernels help the proposed network to find the corresponding symmetry objects.(2)In this paper,features of lower convolutional layers are down-sampled and concatenated with output features of deeper convolutional layers.It improves the feature representation abilities and a better detection performance can be achieved for multi-scale objects.(3)The task of extracting the geometric properties of symmetry objects is accomplished by visualizing and merging activation maps of middle convolutional layer outputs.The supporting regions of rotation symmetry,reflection axes of bilateral symmetry and lattice of translation symmetry can be detected robustly and consistently by merged feature maps.To train the proposed network a symmetry object detection dataset is established.The dataset contains 1308 nature images and annotations include object categories and locations given in a bounding box manner.To verify the effectiveness of the proposed method,a series of experiments are conducted.The results show that the proposed method is capable of detecting symmetry targets under complex scenarios;the merged feature maps detect geometric properties of symmetry objects correctly and accurately.It indicates the effectiveness of the proposed network as well as its sub-network.