Stereo Depth Estimation Algorithms

Introduction

The brain is a fantastic supercomputer with unmatched 3D object recognition and depth estimation. The ocular system has over 120 million rods and 6 million cones which send signals that are funneled into of our visual cortex resulting in our capability to see. The massive amounts of data and processing power of our brain, make it a fascinating phenomenon. Stereo depth is a major part of this machinery and in this paper, we delve into the topic by investigating existing computer algorithms. Depth extraction from stereo estimation boils down to Triangulation. The paper seeks to explain three primary approaches and details surrounding stereo depth extraction.

Background

Computer vision has seen explosive growth during the turn of the century; stereo depth in particular has seen its fair share of research due to its application in panorama mosaic, improved face tracking, facial recognition, autonomous driving, high resolution geometric sampling and many others. The result of this high interest is a plethora of algorithms and papers describing various ways of estimating depth from two images.

A typical stereo algorithm consists of four steps: matching cost computation, cost aggregation, optimization and disparity refinement [1]. Stereo correspondence, the process of matching pixels between stereo images, is essentially an optimization problem. The local or global categories for stereo depth estimation are determined by the way an algorithm solves the stereo correspondence problem.

Methods

"The wide range of modern dense matching algorithms can be assigned to two big categories: local methods, which evaluate the correspondences one point at a time, not considering neighboring points/measures, and global methods where some constraint on the regularity of results are enforced in the estimation"[2]. However, after the advent of the MC-CNN paper[3] a whole host of learning network algorithms have dominated the KITT data set.

Global Method

Energy Minimization

Markov Random Fields have enable a statistical method of building out the disparity method

Other Methods

CNN

MC_CNN is an algorithm that started the movement The latest algorithms have implemented various approaches to improving on the idea of a learned stereo matching. The cost optimization designs, whether global or local, seek to gain an answer of whether a reference pixel is similar or not to the pixel under comparison. The machine learning networks seek to improve on this comparison step by adding apriori knowledge.

Difficult geometry and occlusion of disparate images are considered costs that are to be minimized in previously mentioned implementations. The learned neural nets attempt to be trained with various geometric image snippets with both positive and negative matches to be able to answer the question of similarity with a priori knowledge.

Features are first extracted from input images and then these features are combined together to determine a similarity score

Dynamic Programming

Scan line based traversal with a 2D dp matrix of pixel and disparity