| dc.description.abstract | Stereo matching (or stereo vision) is a field of computer vision that has been getting at-
tention over the last decades, because of its wide range of applications and versatility. It
addresses the problem of 3D reconstruction and depth (disparity) estimation.
The disparity estimation problem can be solved in numerous ways, one them is a block
type solution, in which the input pair passes through stages. The first stage, called the
matching cost computation stage, happens to be one of the most important, because it de-
termines the pairs of pixels that are the most similar between left and right images. The
matching cost function may be a traditional pixel-wise technique, or a deep learning based
function, which is currently the state of the art approach for computing the matching cost.
On this context, the company Woodtech is in need of a stereo vision algorithm that is ca-
pable to solve the disparity estimation problem, using a pair of stereo images, specifically, of
moving truckload images, in an outdoor environment. This memoir contributes in the imple-
mentation of such algorithm, and goes even further by implementing two different matching
cost functions, with the intention of comparing a traditional v/s a deep learning approach.
The algorithms are numerically evaluated with both an indoor and outdoor dataset, pro-
viding a good starting point for knowing each of the algorithms strength and downfalls.
The results prove the superiority of a deep learning approach (v/s the traditional pixel
wise technique chosen) when the images are in outdoor conditions, but also sets a challenge
to make the execution time manageable for a real time application. Nevertheless, the tra-
ditional approach used, showed to be improved when the input images were pre processed,
being almost as good as the deep learning technique, and much less time consuming.
The obtained results are a first step for the company in the field of stereo vision, allowing
them to have a flexible algorithm with two possible matching functions, and a record of the
measured accuracy of each algorithm, which allows them to make a good decision in the
future when it comes to estimating the disparity of their images. | |
