Tube-Link: A Flexible Cross Tube Framework for Universal Video Segmentation

TL;DR

Tube-Link introduces a unified, flexible framework for video segmentation that effectively models cross-tube relationships and achieves significant performance improvements across multiple datasets.

Contribution

The paper presents Tube-Link, a novel framework with tube-level linking and temporal contrastive learning, enabling versatile and efficient video segmentation for various scenarios.

Findings

Achieves nearly 13% improvement on VIPSeg

Improves performance by 4% on KITTI-STEP

Boosts IDOL scores by 3-4% on Youtube-VIS datasets

Abstract

Video segmentation aims to segment and track every pixel in diverse scenarios accurately. In this paper, we present Tube-Link, a versatile framework that addresses multiple core tasks of video segmentation with a unified architecture. Our framework is a near-online approach that takes a short subclip as input and outputs the corresponding spatial-temporal tube masks. To enhance the modeling of cross-tube relationships, we propose an effective way to perform tube-level linking via attention along the queries. In addition, we introduce temporal contrastive learning to instance-wise discriminative features for tube-level association. Our approach offers flexibility and efficiency for both short and long video inputs, as the length of each subclip can be varied according to the needs of datasets or scenarios. Tube-Link outperforms existing specialized architectures by a significant margin on five video segmentation datasets. Specifically, it achieves almost 13% relative improvements on VIPSeg and 4% improvements on KITTI-STEP over the strong baseline Video K-Net. When using a ResNet50 backbone on Youtube-VIS-2019 and 2021, Tube-Link boosts IDOL by 3% and 4%, respectively.