Disentangled self-supervised video camouflaged object detection and salient object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video tasks like VCOD/VSOD are crucial in multimedia.
• Existing methods struggle to effectively utilize both motion and context information.
• Biases between motion and context can limit performance in current models.

Purpose of the Study:

• To propose a novel disentangled perspective for treating motion and context information in VCOD/VSOD.
• To develop a self-supervised approach to reduce annotation costs.
• To improve the performance of VCOD/VSOD tasks by leveraging disentangled information.

Main Methods:

• A new model treats motion and context information separately using ContextNet and MotionNet.
• A self-supervised adaptive frame routing mechanism assigns frames to ContextNet or MotionNet.
• Cross-supervision is employed to train the segmentation networks in a self-supervised manner.

Main Results:

• The proposed self-supervised disentangled model achieves superior performance.
• Consistently outperforms state-of-the-art unsupervised methods on VCOD and VSOD datasets.
• Demonstrates the effectiveness of disentangling motion and context information.

Conclusions:

• The disentangled perspective offers a promising direction for VCOD/VSOD tasks.
• Self-supervised learning significantly reduces annotation requirements.
• The proposed method sets a new benchmark for unsupervised VCOD/VSOD.