Weakly Supervised Camouflaged Object Detection Based on the SAM Model and Mask Guidance
Researchers have developed a new weakly supervised method for camouflaged object detection, a task that involves identifying objects that blend seamlessly with their surroundings. Their approach, called MGNet, uses initial masks from a custom-designed Cascaded Mask Decoder to improve edge predictions and reduce missed detections. To generate training data, they employ BoxSAM, which utilizes the Segment Anything Model (SAM) with bounding-box prompts to create high-quality pseudo-labels. AI
IMPACT This research offers a more efficient approach to camouflaged object detection, potentially reducing the need for extensive manual annotation in computer vision tasks.