Explicit Motion Handling and Interactive Prompting for Video Camouflaged Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Camouflage presents significant challenges for detecting static objects due to their seamless blending with backgrounds.
• Target movement disrupts camouflage, but existing video camouflaged object detection (VCOD) methods struggle with noisy motion estimation or implicit motion modeling in dynamic scenes.

Purpose of the Study:

• To propose a novel Explicit Motion handling and Interactive Prompting (EMIP) framework for VCOD that explicitly leverages motion cues.
• To enhance VCOD performance in complex dynamic scenes by integrating segmentation and motion estimation streams.

Main Methods:

• A two-stream architecture is employed for simultaneous camouflaged segmentation and optical flow estimation.
• Interactive prompting, inspired by visual prompt learning, facilitates cross-stream interactions via learnable 'camouflaged feeder' and 'motion collector' modules.
• Self-supervised learning is used to train the motion stream, and long-term historical information is incorporated as a prompt for temporal consistency.

Main Results:

• The EMIP framework and its long-term variant (EMIP${}^{\dagger }$) establish new state-of-the-art results on popular VCOD benchmarks.
• EMIP${}^{\dagger }$ demonstrates reduced training costs with only 8.5M trainable parameters.
• Comparative evaluations show superior robustness and generalization capabilities compared to other video segmentation models.

Conclusions:

• EMIP effectively addresses limitations of existing VCOD methods by explicitly handling motion cues.
• The proposed framework offers improved accuracy, robustness, and generalization for video camouflaged object detection.
• The approach is computationally efficient, particularly the long-term variant, making it practical for real-world applications.