Priori Information Guided Memory Network for Video Object Segmentation

Abstract

In recent research, memory network-based video object segmentation algorithms have been shown to effectively segment video objects, but these methods did not consider the temporal relationships between frames and did not fully utilize the priori information from previous frames. In order to address this limitation, this paper proposes a Priori information Guided Memory Network(PGNet) for semi-supervised video object segmentation, which effectively utilizes the prior information provided by the previous frame. We use the spatial information from the previous frame to better capture the shape, position, and boundaries of the targets, distinguish the targets and similar object. Meanwhile, when performing spatio-temporal memory reading, it conducts local pixel matching with the prior information provided by the previous frame to maintain consistency of the target objects. The experimental results demonstrate that our algorithm PGNet achieves outstanding performance on the publicly available datasets DAVIS and YouTubeVOS.