Self Supervised Progressive Network for High Performance Video Object Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Self-supervised video object segmentation (VOS) is gaining interest.
• Current single-backbone methods with point-to-point correspondence show limitations in performance.
• Existing approaches often rely on simple pipelines that do not fully leverage video data.

Purpose of the Study:

• To develop a novel self-supervised progressive network (SSPNet) for improved video object segmentation.
• To overcome the performance limitations of existing single-backbone VOS methods.
• To enhance mask propagation and refinement using advanced module designs.

Main Methods:

• Proposed a self-supervised progressive network (SSPNet) comprising a memory retrieval module (MRM) and a collaborative refinement module (CRM).
• MRM utilizes pixel-level and frame-level similarity learning for point-to-point correspondence and coarse mask generation.
• CRM employs cycle consistency region tracking for refining masks by aggregating reference and query information.
• Introduced two novel mask-generation strategies to incorporate semantic knowledge from unlabeled data.

Main Results:

• SSPNet demonstrated superior performance compared to state-of-the-art self-supervised VOS methods.
• The proposed method significantly narrowed the performance gap with fully supervised VOS techniques.
• Experiments on DAVIS-17, YouTube-VOS, and SegTrack v2 validated the effectiveness of SSPNet.

Conclusions:

• SSPNet offers a more effective approach to self-supervised video object segmentation.
• The combination of MRM and CRM enables robust mask propagation and refinement.
• The method shows promise for advancing self-supervised learning in VOS tasks.