G. Mihalev, S. Yordanov, H. Stoycheva, Application of Segmentation Algorithms in the Reconstruction of a 3D Scene from Images

Key Words: 3D reconstruction; Structure From Motion (SFM); Image segmentation; Robotics.
Abstract. The technique of 3D reconstruction of real scenes from images represents a significant advancement in visual technologies, offering the potential for the development of numerous new applications in various engineering fields. In general, 3D reconstruction from images is a computer process that utilizes a set of two-dimensional images captured from different angles and positions to create a three-dimensional model of the scene or object. By analyzing features such as key points, comparing their positions, and applying geometric principles, this process transforms the images into three-dimensional representations with spatial information. The complexity of 3D scene reconstruction from images arises from the need to detect points of interest, match and track them across different images, as well as analyze and merge multiple visual data from different perspectives. This article proposes the use of simpler image processing algorithms to obtain more points of interest, providing a more accurate 3D model without overburdening the reconstruction process with complex computational procedures. The investigated methods implement image segmentation through the first and second derivatives of the image intensity function to determine and separate edges and boundaries in the image. A practical implementation of 3D reconstruction of real scenes from images in the Matlab environment is presented. In the implementation, algorithms based on the first and second derivatives of the image intensity function are used to achieve a denser 3D model by segmenting images to detect edges and boundaries. The investigated algorithms are implemented in the stage of discovering properties in the image. Some of the most well-known algorithms have been tested and compared with the use of the standard SURF algorithm. The algorithms from this class have significantly improved results, with the density of the 3D model increased several times compared to the number of generated points, and in some cases, the execution time remains the same relative to the percentage ratio of the generated 3D points.