Nowhere to Disguise: Spot Camouflaged Objects via Saliency Attribute Transfer

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Salient Object Detection (SOD) and Camouflaged Object Detection (COD) are crucial object segmentation tasks.
• While seemingly contradictory, SOD and COD share underlying principles related to object recognition and contextual understanding.

Purpose of the Study:

• To explore the relationship between SOD and COD.
• To adapt successful SOD models for COD tasks to minimize design expenses.
• To develop a novel framework for decoupling and transferring contextual information between SOD and COD.

Main Methods:

• A novel decoupling framework with triple measure constraints was designed to separate object semantic representations and context attributes.
• An attribute transfer network was introduced to transfer saliency context attributes to camouflaged images.
• Weakly camouflaged images were generated to bridge the context attribute gap.

Main Results:

• The proposed method successfully transfers context attributes from SOD to COD datasets.
• SOD models adapted using this method show improved performance on COD tasks.
• Experiments on three widely-used COD datasets validate the effectiveness of the approach.

Conclusions:

• The study establishes a strong connection between SOD and COD, demonstrating their shared reliance on semantic and contextual information.
• The developed attribute transfer method enables the effective adaptation of SOD models for COD.
• This research offers a cost-effective solution for camouflaged object detection by leveraging existing SOD models.