Profile Facial Expression Recognition And Object Detection Using A Spherical Camera

A spherical camera can have a full field of view.It is suitable for the tasks which need the visual information of the surrounding environments.In this paper,the following two applications are developed by exploiting the characteristics of a spherical camera.1.Profile facial expression recognition by a head-mounted spherical camera Facial expression recognition is one of the most intuitive expressions of human emotion.If the user's facial expression can be recognized through VR and other headgear,the user's most real experience can be obtained,and how to recognize the facial expression when the user is free to move is a new problem to be solved.In this paper,we propose a novel method of face expression recognition via the profile image captured by a head-mounted spherical camera.First,we created a fish eye profile image data-set.Then,we trained a 3D convolutional neural network proposed by DU Tran [1] for dynamic expression recognition.As shown in the experimental results,we achieved as high as 72.2% face expression recognition rate.2.Surrounding objects detection using convolutional neural networks trained by perspective imagesConvolutional Neural Networks(CNNs)have achieved a remarkable success in image processing and computer vision tasks.However,most popular CNNs are developed for and trained by perspective images.It is known that training a deep CNN is a high-cost process.Can we use CNNs trained by perspective images to process omnidirectional images? In this paper we propose a method of detecting objects from an omnidirectional image based on circular-perspective representation,called CPR,which is also called a panoramic image in some previous research.A circular-perspective representation image is a panoramic image which horizontal axis is azimuth angle along the circle of a cylinder and its vertical axis obeys the traditional perspective projection.Since for a horizontally small region in CPR can be seen as an approximation of a perspective image patch,the conventional Convolutional Neural Networks trained by perspective images can be applied to such a CPR image directly.As an example,an experimental result on objection detection is given to show the effectiveness of the proposed method.