Convoluciones 3D en paralelo
Loading...
Identifiers
Publication date
Authors
Advisors
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
En este Trabajo de Fin de Máster se construye y optimiza un algoritmo de convolución 3D que se encarga de procesar nubes de puntos LiDAR aéreo que permita identificar objetos similares según el patrón de entrada dado. Concretamente, se utiliza un conjunto de datos correspondiente a zonas residenciales y, para evaluar el funcionamiento del algoritmo, se marca como objetivo la identificación de edificaciones. Tras la construcción del algoritmo de convolución y el ensayo con diferentes estructuras de representación de la nube de puntos, se aplican optimizaciones de paralelización en memoria compartida tratando de mantener un consumo de memoria reducido. Tras aplicar las técnicas consideradas, se consigue reducir el tiempo de ejecución de la convolución en un 98%, utilizando para ello 64 núcleos en 2 procesadores con OpenMP.
In this Master Thesis, a 3D convolution algorithm is built and optimized to process aerial LiDAR point clouds in order to identify similar objects according to the given input pattern. Specifically, a dataset corresponding to residential areas is used and, in order to evaluate the performance of the algorithm, the identification of buildings is set as a target. Following the construction of the convolution algorithm and experimentation with different point cloud representation structures, optimizations for shared memory parallelization are applied trying to achieve a low memory consumption. After applying the considered techniques, the execution time of the optimized part is reduced by 98%, using 64 cores in 2 processors with OpenMP.
In this Master Thesis, a 3D convolution algorithm is built and optimized to process aerial LiDAR point clouds in order to identify similar objects according to the given input pattern. Specifically, a dataset corresponding to residential areas is used and, in order to evaluate the performance of the algorithm, the identification of buildings is set as a target. Following the construction of the convolution algorithm and experimentation with different point cloud representation structures, optimizations for shared memory parallelization are applied trying to achieve a low memory consumption. After applying the considered techniques, the execution time of the optimized part is reduced by 98%, using 64 cores in 2 processors with OpenMP.
Description
Bibliographic citation
Relation
Has part
Has version
Is based on
Is part of
Is referenced by
Is version of
Requires
Sponsors
Rights
Attribution-NonCommercial-ShareAlike 4.0 International








