3D Model Retrieval And Pose Estimation Using Feature Learning

With the occurrence and continuous development of the Internet,the forms of information that the people manage to acquire have witnessed a great change—the forms used to be the ones of text or numbers,and are currently the ones of images,audios,videos,etc.Nowadays multi-media are in great need of 3D models and requiring a higher speed as well as a better quality in 3D modeling,so it is quick and of great efficiency to modify those related and existed 3D models via 3D model retrieval.3D model retrieval takes 3D model features as index.Both the choices of 3D model features and the ways of optimizing 3D model features are equally essential problems to be faced with and solve while we develop a 3D model retrieval system.It is proposed in the paper that a new method of optimizing 3D model features using Kernel Fisher Discriminant Analysis is intended to improve the result of 3D model retrieval.The main idea of the paper is to map the linearly indivisible samples to a high-dimension space which meets the condition of Mercer at first and then project the original 3D model features to the subspace using Linear Discriminant Analysis in the high-dimension space.In this way the discriminant 3D model features used in 3D model retrieval can be obtained to separate class better.As is shown in the experiment,fast in speed,not only is Kernel Fisher Discriminant Analysis capable of optimizing 3D model features in seconds,but improving the query quality by 15%in average as well.Pose estimation problem is the one that estimates the pose parameters such as the directions and scales of a 3D model after detecting the position of a target object in a 2D image.In order to decline the amounts of calculation and the effects which man-made features bring about,a new method is adopted which takes advantage of Convolutional Neural Network to estimate pose based on a single 3D model depth map.Its main idea is to train an appropriate Convolutional Neural Network architecture with the help of supervised learning at the beginning,and then to estimate the pose parameters after extracting the features of 3D model depth maps.During the processing man-made features are absent,which reduces some subjectivity in the estimation.