A Content Enhancement Framework for Multi-Projector Systems

Abstract

Projectors are a convenient technology for displaying content on large, abnormal, or temporary surfaces where mounting other forms of light emitting devices is too impractical or too expensive. Common uses of projectors include movie cinemas, concert halls, 3D model colourization, planetariums, etc. Many of these applications require multiple projectors to either cover the entire display surface, like planetariums, or to achieve the require brightness, like outdoor projection.

Aligning the content between projectors is typically required to ensure that overlapping regions between projectors display the same content. Naive approaches of aligning content treat the relationship between the content and a projector independently of all other projectors in the configuration. Aligning content can limit the quality of the superimposed image as high frequency signals are often degraded during the alignment process. Previous works have shown it is possible to improve the perceptual quality of the aligned content by giving each content-to-projector transformation prior knowledge of all projectors in the configuration. However, these works either make theoretical assumptions, require special hardware, severely limit the types of applications their systems work on, or only use qualitative analysis to evaluate their system's performance.

In this work, a framework capable of simulating a multi-projector configuration for any number of projectors on a flat surface is proposed. A method of comparing the ideal content with the projected content is developed using the proposed simulation in conjunction with an existing image comparison technique. Different system setups are tested for a two projector configuration. The quality of each configuration is measured using the developed comparison metric across a dataset of natural images. Finally, the proposed framework is used to train three different models, in an end-to-end fashion, that are capable of improving the perceptual quality of the superimposed image.

The first two models are parametric and content independent, while the third model is non-parametric and content dependent. The first model directly integrates with existing interpolation methods used during the content-to-projector alignment. The second model applies a post transformation filtering operation using a set of learned linear convolutional kernels. The third model directly optimizes the projected images to improve the perceptual quality of the superimposed image.