Predição de mapas de profundidades a partir de imagens monoculares por meio de redes neurais sem peso

Abstract

Depth estimation – taking one or more images from a scene and estimating a depth map, which determines distances between the observer and points taken from various object surfaces – is a central problem in computer vision. It's not surprising, then, that the approach of stereo correspondence, traditionally applied to this problem, is one of the most intensively studied topics in the field. Stereo correspondence systems estimate depths from binocular features – more specifically, the difference between the positions of each point in a pair of images. Besides this purely geometrical information, images contains many monocular features – such as texture variations and gradients, focus, color patterns and reflection – which can be explored to derive depth estimates. For this, however, a certain amount of a priori knowledge must be gathered, since there is an inherent ambiguity between an image's characteristics and depth variations. Through his research on machine-learning systems based on Markov Random Fields (MRF's), Ashutosh Saxena proved that it is possible to estimate very accurate depth maps from a single monocular image. His approach, however, lacks biological plausibility, since there is no known theoretical correspondence between MRF's and the human brain's neural networks. Motivated by past successes in applying Weightless Neural Networks (WNN's) to computer vision problems, in this paper we investigate the effectiveness of applying WNN's to the problem of depth map estimation. With this, we hope to achieve performance improvements over Saxena's MRF-based approach, and develop a more useful architecture for evaluating hypotheses about visual information processing in the human cortex.