Deep Features for Categorization of Heritage Images Towards 3D Reconstruction

摘要

In this paper, we propose a framework to categorize heritage site images based on unsupervised clustering for 3D reconstruction. Modeling a generalized classification model for crowdsourced data is challenging as the data is invariant and high dimensional. Handcrafted features find challenges in differentiating such data. To address this, we propose a model to cluster the data using deep features. We model Inception based Variational Autoencoder(IVAE) for extracting deep 8 features from crowdsourced data in an unsupervised manner. Towards refining deep features we propose a new loss function, variant of Variational Autoencoder (VAE) loss function and model IVAE accordingly. The extracted features are represented in latent space and clustered using clustering algorithms like K-Means, Mini-batch-K-Means. The experimental results obtained outperforms the state of art methods in terms of clustering accuracy.