MRCNet: Motion Reasoning Chain for Cross Modal Video Camouflaged Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Video camouflaged object detection (VCOD) traditionally relies on visual cues and inter-frame motion.
• High object-environment similarity and noisy motion (e.g., camera shake) limit traditional VCOD performance.

Purpose of the Study:

• To propose a novel cross-modal VCOD framework, the Motion Reasoning Chain Network (MRCNet).
• To emulate human motion reasoning for enhanced detection of camouflaged objects in videos.

Main Methods:

• Utilizing a generative sampling strategy with multimodal large language models (MLLMs) to bridge implicit knowledge and explicit object attributes.
• Developing motion representation learning driven by a motion reasoning chain, incorporating hierarchical de-biased motion prototype learning.
• Employing cross-modal prompt learning to integrate de-biased concept prototypes into visual representations for improved comprehension.

Main Results:

• MRCNet achieves state-of-the-art performance on general and spatiotemporal consistency metrics across three datasets.
• The framework effectively establishes a motion reasoning chain tailored for VCOD.
• Hierarchical de-biased motion prototype learning mitigates MLLM hallucinations and boosts motion perception.

Conclusions:

• MRCNet offers a significant advancement in video camouflaged object detection by integrating cross-modal reasoning.
• The proposed motion reasoning approach enhances both the accuracy and temporal consistency of detecting challenging camouflaged objects.