SVNet: Supervoxel Network for Video Oversegmentation

doi:10.1145/3650400.3650460

科研成果详情

题名	SVNet: Supervoxel Network for Video Oversegmentation
作者	Qi，Yijie 1; Yang，Baorong 1; Zhang，Wenjing 2; Dong，Xiao 3
发表日期	2023-10-20
会议录名称	ACM International Conference Proceeding Series
页码	368-376
摘要	Supervoxel segmentation for video is a video pre-processing technique that groups voxels with similar spatiotemporal features into supervoxels, effectively reducing the number of elemental voxels for downstream computer vision applications, or applications in other fields, e.g., hardware design or video visualization. Existing methods for video supervoxel generation primarily rely on traditional techniques, such as graph theory, mean shift, and clustering, which have yielded promising results. Recent deep learning-based methods mainly worked on object segmentation or semantic segmentation from videos, paying less attention to video oversegmentation. However, the quality of supervoxels directly affects the results of subsequent tasks. In this paper, we introduce a novel approach SVNet which enables direct end-to-end segmentation of voxels into supervoxels from a deep iterative clustering network. The process begins by utilizing the spatiotemporal features learned by the deep neural network to construct a soft association map between voxels and supervoxels. Subsequently, through an iterative update process, the features of supervoxels and the soft association map between voxels and supervoxels are continually refined to enhance the accuracy of voxels segmentation into supervoxels with the supervised loss using the reconstructed video features and labels. We evaluate our method and the representative supervoxel algorithms for their capability on the performance of video segmentation. Experiments show that our SVNet excels particularly in terms of the BRD metric, and its accuracy is roughly on par with the compared methods.
关键词	Clustering Spatio-temporal learning Supervoxels Video segmentation
DOI	10.1145/3650400.3650460
URL	查看来源
语种	英语English
Scopus入藏号	2-s2.0-85191488597
引用统计
文献类型	会议论文
条目标识符	https://repository.uic.edu.cn/handle/39GCC9TT/11560
专题	北师香港浸会大学
通讯作者	Yang，Baorong
作者单位	1.College of Computer Engineering,Jimei University,Xiamen,361021,China 2.School of Mathematics and Computer Science,Zhejiang A&f University,Hangzhou,311300,China 3.Guangdong Provincial Key Laboratory Irads,BNU-HKBU United International College,Zhuhai,519087,China
推荐引用方式 GB/T 7714	Qi，Yijie,Yang，Baorong,Zhang，Wenjinget al. SVNet: Supervoxel Network for Video Oversegmentation[C], 2023: 368-376.