Weakly supervised video object segmentation initialized with referring expression

Abstract

With the aid of one manually annotated frame, One-Shot Video Object Segmentation (OSVOS) uses a CNN architecture to tackle the problem of semi-supervised video object segmentation (VOS). However, annotating a pixel-level segmentation mask is expensive and time-consuming. To alleviate the problem, we explore a language interactive way of initializing semi-supervised VOS and run the semi-supervised methods into a weakly supervised mode. Our contributions are two folds: (i) we propose a variant of OSVOS initialized with referring expressions (REVOS), which locates a target object by maximizing the matching score between all the candidates and the referring expression; (ii) segmentation performance of semi-supervised VOS methods varies dramatically when selecting different frames for annotation. We present a strategy of the best annotation frame selection by using image similarity measurement. Meanwhile, we first to propose a multiple frame annotation selection strategy for initialization of semi-supervised VOS with more than one annotated frames. Finally we evaluate our method on DAVIS-2016 dataset, and experimental results show that REVOS achieves similar performance (79.94% measured by average IoU) compared with OSVOS (80.1%). Although current REVOS implementation is specific to the method of one-shot video object segmentation, it can be more widely applicable to other semi-supervised VOS methods.