Methods for Fast and Accurate Depth Estimation from Sparse Light Fields

Abstract

Depth estimation is a crucial challenge for light-field technology as it forms the foundation for multiple important applications, such as 3D scene reconstruction and image-based rendering. While significant progress has been made in developing new depth estimation methods, most of them have focused on accuracy, while aspects of practical use, such as computational costs and scalability, have rarely been addressed. Depth estimation methods that offer a better balance between computational time and reconstruction accuracy are especially needed.

This thesis presents a fast and accurate method for dense depth estimation from sparse, wide-baseline light fields obtained by camera arrays. In the proposed method, the input multi-perspective images are over-segmented into non-overlapping compact superpixels modelled as planar image patches. Superpixels are used as basic primitives for initial depth estimation and subsequent refinement. The resulting sparse superpixel-based representation allows for the targeted reduction in both memory and computational costs. At the same time, the proposed method reconstructs highly accurate dense depth maps with precise object contours.

Furthermore, to improve the accuracy of the superpixel segmentation, an efficient method for content-adaptive superpixel generation is proposed. Instead of adapting the size and distribution of superpixels to the image content, the method introduces an image warping transform that redistributes the image content and makes it more suitable for subsequent segmentation into regular superpixels. In turn, regular superpixels in the warped image induce content-adaptive superpixels in the original image with improved segmentation accuracy. A parallelised iterative coarse-to-fine optimisation procedure has been proposed for efficient computation of the warping transform.

Experiments with efficient GPU implementations of both methods validate their feasibility in achieving an improved trade-off between quality and computational requirements, which is instrumental in practical light-field applications.