Invariant Object Recognition using Circular Pairwise Convolutional Networks Choon Hui Teo 1 and Yong Haur Tay 2 1 National ICT Australia Ltd. and Research School of Information Sciences and Engineering Australian National University choonhui.teo@rsise.anu.edu.au 2 Faculty of Information and Communication Technology Universiti Tunku Abdul Rahman tayyh@mail.utar.edu.my Abstract. Invariant object recognition has been one of the most re- warding are of research in computer vision as there are many applications need the capability of recognizing objects of interest in various environ- ments. However, there is no single technique which claims to achieve the goal in all possible conditions and domains. Out of many techniques, convolutional network has proved to be a good candidate in this area. Given large numbers of training samples of objects under various varia- tion aspects such as lighting, pose, background, etc., convolutional net- work can learn to extract invariant features by itself. This comes with the price of lengthy training time. Hence, we propose a circular pairwise classification technique to shorten the training time. We compared the recognition accuracy and training time complexity between our approach and a benchmark generic object recognizer LeNet7 which is a monolithic convolutional network. 1 Introduction Invariant Object Recognition is an object recognition approach which deals with recognizing an interested object in an image under some variation aspects. Pos- sible variation aspects are translation, scaling, shearing, rotation, distortion, illuminations, angle of view (pose), occlusion, noisy (jittered and/or cluttered) background, changes in object colors and texture and etc. An object can be defined as a set of geometrical, structural or pictorial fea- tures (or characteristics) which exhibit large inter-class variance while keeping intra-class variance small at the same time. Such features could be geometri- cal shape, number of interconnected cylinders with certain relationship, color histogram, texture, height-width ratio, etc. In invariant object recognition, the first challenge would be defining a set of detectable invariant features for each class of interest such that objects can be recognized under many variation aspects. In order to devise such ideal ob- ject representation, many research rooted in statistics (e.g. K-Nearest Neighbor),