Hierarchical Co-Attention Propagation Network for Zero-Shot Video Object Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot video object segmentation (ZS-VOS) aims to segment objects without prior knowledge.
• Existing ZS-VOS methods face challenges in distinguishing foreground/background and tracking objects in complex scenarios.
• Overreliance on optical flow estimation is a limitation in current approaches.

Purpose of the Study:

• To propose a novel encoder-decoder-based Hierarchical Co-attention Propagation Network (HCPN) for ZS-VOS.
• To address the limitations of existing methods in foreground/background distinction and object tracking.
• To develop a model that reduces overreliance on optical flow estimation.

Main Methods:

• The proposed HCPN utilizes parallel co-attention modules (PCM) and cross co-attention modules (CCM).
• PCM captures common foreground regions between appearance and motion features.
• CCM fuses cross-modal motion features, and the model is progressively trained for hierarchical spatio-temporal feature propagation.

Main Results:

• The HCPN demonstrates superior performance compared to all previous methods on public benchmarks.
• Experimental results validate the effectiveness of the proposed HCPN for ZS-VOS tasks.
• The model successfully tracks and segments objects in complex video sequences.

Conclusions:

• The HCPN offers a significant advancement in zero-shot video object segmentation.
• The proposed co-attention mechanism effectively handles appearance and motion information for robust segmentation.
• The method provides a strong baseline for future research in ZS-VOS.