Apparatus for performing cylindrical convolutional neural network operation and apparatus and method for object recognition and viewpoint estimation using the same

An apparatus for object recognition and viewpoint estimation, comprising:

• a preprocessor unit for receiving an input image including at least one object and extracting an object region feature map that is a region of each of the at least one object in the input image;
• a cylindrical kernel unit, including a cylindrical kernel having a structure in which a plurality of weight columns corresponding to a pre-designated angular unit interval are periodically repeated in units of 360 degrees, and performing a convolution operation on the number of weight columns corresponding to an angular range of a pre-designated size among the plurality of weight columns of the cylindrical kernel and the object region feature map, thereby extracting a plurality of view-specific feature vectors according to each angular unit interval;
• a score map acquisition unit, extracting a reliability score according to object class for each of the plurality of view-specific feature vectors by using an artificial neural network in which a plurality of object classes and a plurality of patterns of viewpoints for each object have been learned, thereby obtaining a plurality of score maps;
• and
• an object and viewpoint determining unit, analyzing the plurality of score maps obtained according to viewpoint corresponding to each of the plurality of view-specific feature vectors, thereby determining the class of the object and the viewpoint included in the object region.

A method for object recognition and viewpoint estimation, the method comprising the steps of:

• receiving an input image including at least one object, and extracting an object region feature map that is a region of each of the at least one object in the input image;
• including a cylindrical kernel having a structure in which a plurality of weight columns corresponding to a pre-designated angular unit interval are periodically repeated in units of 360 degrees, and performing a convolution operation on the number of weight columns corresponding to an angular range of a pre-designated size among the plurality of weight columns of the cylindrical kernel and the object region feature map, thereby extracting a plurality of view-specific feature vectors according to each angular unit interval;
• extracting a reliability score according to object class for each of the plurality of view-specific feature vectors by using an artificial neural network in which a plurality of object classes and a plurality of patterns of viewpoints for each object have been learned, thereby obtaining a plurality of score maps;
• and
• analyzing the plurality of score maps obtained according to viewpoint corresponding to each of the plurality of feature vectors according to viewpoint, thereby determining the class of the object and the viewpoint included in the object region.