Guided Filters for Depth Image Enhancement

Saad, Mohammad

Permalink

https://hdl.handle.net/2142/97838 Copy

Description

Title

Guided Filters for Depth Image Enhancement

Author(s)

Saad, Mohammad

Contributor(s)

Do, Minh N.

Issue Date

2016-12

Keyword(s)

• Depth Enhancement
• Depth Sensor
• Guided Filtering
• Depth Refinement
• Temporal Consistency
• Video Processing
• Image Inpainting

Abstract

This thesis proposes an approach utilizing guided techniques to refine depth images. Given a depth image and a color image of the same resolution, we can utilize the color image as a guide to improve the accuracy of the depth image, by smoothing out edges and removing holes as much as possible. This is done utilizing a guided filter, which solves an optimization problem relating the depth and color image to smooth and refine the depth image. These guided filters are linear-time and much faster than other state-of-the-art methods, while producing comparable results. We also integrate an existing guided inpainting model, further removing holes and improving the depth map. In this thesis, we show the application of guided filters to the depth refinement problem, utilize a guided inpainting model to fill in any holes that may arise in the depth image, as well as extend the filter out to the temporal domain to handle temporal flickering. This is done via an extension of existing optical-flow methods to compute a weighted average of the previous and next neighbors. We also demonstrate a few experimental results on real-time video to show that this method has viability in consumer depth applications. We demonstrate results on both datasets and real video to show the accuracy of our method.

Type of Resource

text

Language

en

Permalink

http://hdl.handle.net/2142/97838

Owning Collections