Attention-Guided Mask Propagation for Video Object Segmentation

Abstract

Video Object Segmentation (VOS) is a pixel-level classification task aiming to separate target objects from the background portion of a video. Temporally consistent pixel-level partition of video sequences has applications across several domains such as object tracking, video summarization, video compression and editing, human-computer interaction, and autonomous vehicles. In this document we focus on the task of VOS in a semi-supervised setting, that is where the first frame ground truth is available and defines the target for the rest of the video sequence. We will analyze different state-of-the-art techniques for VOS with particular focus on their strengths and weaknesses. Then, we will introduce our design for a deep learning-based approach that achieves competitive results on available public benchmarks. We will present a detailed comparison of the performances in relation to the current state-of-the-art and investigate the effect of each component of our model with extensive ablation studies. Finally, we will suggest possible future research paths both for the VOS field in general and more specifically for our architecture.