Multi-View 3D Object Retrieval using Autoencoder & Deep Embedding Network

Due to the rapid development of 3D object, there are many application that need to be retrieved 3D object which is very useful in computer vision based problems. A collection of various aspect of a particular object's 2D images are referred as multi- view 3D object. Rather than using hand-crafted features extraction techniques, two unique feature extractor, deep embedding network and autoencoder are used to learn the features and measured the correlation between trained features and query image's features by using Euclidean distance for retrieving images. This paper focus on the mechanism for retrieving 3D object from both seen and unseen categories multi-view 2D images. Both supervised and unsupervised models are trained on a set of datasets comprising a specific set of classes which is called seen classes or categories and the retrieval operation is performed on set of other classes which are not present in trained dataset is the unseen category set along with seen categories. Deep embedding network is mutually supervised by classification loss and triplet loss to optimize the intra-class variation for image retrieval. Finally, to compare image retrieval performance accuracy, various assessment criteria are used to prove the effectiveness of using deep embedding network & autoencoder as feature extractor and compare computation time and space complexity. Here, PCA and t-SNE are also used to analyze the retrieval factor of these networks. After implementing our proposed methods, deep embedding network achieved 97% accuracy and autoencoder achieved 96% accuracy to retrieve multi-view 2D images using RGB-D dataset on seen categories. For the unseen categories retrieval accuracy of deep embedding network and autoencoder it is 92% and 95% accordingly.