Semi-Automated Stereo Image Patches Generation and Labeling Method Based on Perspective Transformations

Diego P. Durante; Ramiro Verrastro; Juan Carlos Gómez; Claudio Verrastro

Diego P. Durante Univ. Tecnologica Nacional Fac. Regional Bs. As.
Ramiro Verrastro Univ. Tecnologica Nacional Fac. Regional Bs. As.
Juan Carlos Gómez Univ. Tecnologica Nacional Fac. Regional Bs. As. - Centro de Electrónica e Informática, Instituto Nacional de Tecnología Industrial
Claudio Verrastro Univ. Tecnologica Nacional Fac. Regional Bs. As. - Comisión Nacional de Energía Atómica

Palabras clave: Computer Vision, Machine Learning, Wide Baseline Stereo, Labeling Tool, Siamese Convolutional Neural Networks

Resumen

In computer vision, Wide Baseline Stereo (WxBS) refers to Vision System configurations on which their images come from cameras with non parallel and widely separated views.

One common task in reconstruction algorithms of WxBS consists of subvididing the stereo images in multiple image patches and then associate
homologous patches between homologous images. Multiple approaches
can be used to associate homologous patches. To train and test supervised
learning algorithms for this tasks, a labeled dataset is required.

In this work, a semi-automated method to generate patches and their
labels from WxBS images is presented. It allows to calculate thousands
of positive and negative pairs of patches with a score of correspondence
between a pair of potentially homologous image patches. This method
largely solves the problems of traditional approach, which requires a lot of
hand labeled work and time. To apply the method, images from different
viewpoints of objects with planar faces and their corner locations are
required.