Learning Long-term Structural Dependencies for Video Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current video salient object detection (VSOD) methods often overlook inter-frame structural dependencies by focusing on global frame-level or grid-based temporal information.
• Existing approaches fail to capture the complex relationships between object structures across different frames in a video.

Purpose of the Study:

• To propose a novel method for video salient object detection (VSOD) that learns long-term structural dependencies.
• To enhance the understanding of inter-frame structural relationships for more accurate salient object identification.

Main Methods:

• A structure-evolving graph convolutional network (GCN) is proposed, constructing a graph based on spatio-temporal structural similarity using supervoxel segmentation.
• Convolutional operations on the graph infer inter-frame structural dependencies of salient objects.
• An adaptive graph pooling mechanism dynamically merges similar nodes to evolve the graph structure, pruning redundant connections and improving hierarchical representations.

Main Results:

• The proposed method significantly outperforms state-of-the-art methods on six public datasets for video salient object detection.
• The adaptive graph pooling technique demonstrably improves the segmentation accuracy of the underlying supervoxel algorithm.

Conclusions:

• The structure-evolving GCN effectively models long-term structural dependencies for improved VSOD.
• Adaptive graph pooling is a valuable technique for enhancing graph-based methods in computer vision tasks, particularly for dynamic objects.