Skip to Main content Skip to Navigation

Generic heuristics on GPU to superpixel segmentation and application to optical flow estimation

Abstract : Finding clusters in point clouds and matching graphs to graphs are recurrent tasks in computer science domain, data analysis, image processing, that are most often modeled as NP-hard optimization problems. With the development and accessibility of cheap multiprocessors, acceleration of the heuristic procedures for these tasks becomes possible and necessary. We propose parallel implantation on GPU (graphics processing unit) system for some generic algorithms applied here to image superpixel segmentation and image optical flow problem. The aim is to provide generic algorithms based on standard decentralized data structures to be easy to improve and customized on many optimization problems and parallel platforms.The proposed parallel algorithm implementations include classical k-means algorithm and application of minimum spanning forest computation for super-pixel segmentation. They include also a parallel local search procedure, and a population-based memetic algorithm applied to optical flow estimation based on superpixel matching. While data operations fully exploit GPU, the memetic algorithm operates like a coalition of processes executed in parallel on the multi-core CPU and requesting GPU resources. Images are point clouds in 3D Euclidean space (space-gray value domain), and are also graphs to which are assigned processor grids. GPU kernels execute parallel transformations under CPU control whose limited role only consists in stopping criteria evaluation or sequencing transformations.The presented contribution contains two main parts. Firstly, we present tools for superpixel segmentation. A parallel implementation of the k-means algorithm is presented with application to 3D data. It is based on a cellular grid subdivision of 3D space that allows closest point findings in constant optimal time for bounded distributions. We present an application of the parallel Boruvka minimum spanning tree algorithm to compute watershed minimum spanning forest. Secondly, based on the generated superpixels and segmentation, we derive parallel optimization procedures for optical flow estimation with edge aware filtering. The method includes construction and improvement heuristics, as winner-take-all and parallel local search, and their embedding into a population-based metaheuristic framework. The algorithms are presented and evaluated in comparison to state-of-the-art algorithms.
Document type :
Complete list of metadatas

Cited literature [142 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Monday, July 20, 2020 - 2:59:12 PM
Last modification on : Tuesday, July 21, 2020 - 4:22:48 AM


Version validated by the jury (STAR)


  • HAL Id : tel-02902962, version 1


Abdelkhalek Mansouri. Generic heuristics on GPU to superpixel segmentation and application to optical flow estimation. Other [cs.OH]. Université Bourgogne Franche-Comté, 2020. English. ⟨NNT : 2020UBFCA012⟩. ⟨tel-02902962⟩



Record views


Files downloads