Compressive Multi-View Rendering: Problem Formulation and Resolution

Abstract

Compressive sensing (CS) is a sampling theory that aims to reconstruct signals from fewer measurements than is done in the classical Nyquist-Shannon sampling scheme. Aside from image coding, CS has been recently leveraged successfully in several rendering acceleration tasks. In this work, we generalize the recent success of CS in 3D rendering to a multi-view setup. We formulate the problem as a joint reconstruction of partially rendered views using the CS. A dictionary learning approach is used to leverage signal sparsity condition for the multi-view reconstruction. The reconstruction process was guided by the depth of the scene, which constitutes valuable and computationally efficient information on the geometry of the 3D scene. Preliminary results showed a significant improvement in both the synthetic image quality and rendering time.