【摘要】 场景解析是计算机视觉算法的基本任务之一。随着深度传感器的发展,获取RGBD场景数据已经变的较为方便。针对RGBD场景解析,提出了一种基于CUDA架构的高速算法。该方法是在gSLICr算法基础上,引入带空间平滑约束的深度信息,结合边缘支持,形成BSD-gSLICr算法,对场景进行超高速过分割。结合多特征的带权Felzen-Hutten算法,将过分割块构成的无向图进行多层级的图割归并,实现高速场景解析。实验结果表明提出的算法在保证效果和鲁棒性的基础上,运行速度可以达到实时要求。更多还原

【Abstract】 Scene parsing is a fundamental task in computer vision.With the development of depth sensor,the RGBD scene data have become more inexpensive to obtain.We propose a high speed algorithm base on CUDA architecture which is focusing on RGBD scene parsing.We bring depth information with spatial smoothing constraints as well as boundary support into gSLICr algorithm to oversegment the scene.Then a hierarchical weighted Felzen-Hutten algorithm is used to optimize the result,considering the scene oversegmentation as an undirected graph.The experiments shows that proposed algorithm can meet the realtime requirement with a highly considerable effect and robustness.更多还原