Efficient Temporal Denoising for Improved Depth Map Applications

Pengzhi Li, Zhiheng Li

Abstract: Depth estimation involves acquiring three-dimensional information from images, which has numerous applications in downstream tasks. Although several effective monocular depth estimation algorithms have been developed, directly applying frame-by-frame depth estimation can result in flickering, which hinders many video-related applications. Previous video-based approaches have primarily been post-processing methods that utilize spatial information about camera poses to reduce flicker, but they come with a considerable computational cost. In this paper, we introduce the concept of depth map noise to better understand flicker in depth maps and propose a depth noise smoothing network to eliminate visual flicker in depth maps. Our approach can be applied to different depth estimation models and run in real-time for screen-based applications, such as video bokeh.