Recovering Multi-frame Incomplete Path Tracing Images Using Tensor Completion

Abstract

This paper presents a novel method to recover missing pixels in multi-frame incomplete path tracing images using tensor completion. Matrix completion and compressed sensing provide screen space solutions to recover missing pixels from sparsely rendered path tracing images, providing efficient previsualization. Path tracing is often used to generate an animated sequence of images, which then requires an efficient previsualization that is coherent across multiple frames to avoid temporal noise in the final video. We present a novel numerical method to construct a tensor structure using multiple path tracing images followed by tensor completion, which coherently recovers missing pixels avoiding temporal noise. Our numerical method avoids complex procedures when building a low rank tensor for tensor completion, which previously required complex initialization and similar patch finding steps across nearby frames. Our method shows promising results and outperforms recent matrix completion based methods in both visual quality and performance.