Year: 2025
Type: article
Abstract: The domain of Camouflaged Object Detection (COD) focuses on the intricate task of discerning objects that are visually integrated into their surroundings. The inherent complexity of COD stems from the seamless blending of camouflaged objects with their environment and the indistinct boundaries these objects often possess. To address these challenges, we have meticulously assembled a comprehensive dataset, designated as COD10K. This dataset encompasses over 10,000 images featuring objects camouflaged across diverse natural landscapes, covering 78 distinct object categories. Each image within COD10K is extensively annotated with category labels, bounding boxes, soft groupings, item instances, and levels of camouflage, laying a robust foundation for advancing research in several vision-based tasks, including image segmentation, object localization, and alpha-matting. Moreover, we introduce the Search Identification Network (SINet), a novel and efficient model tailored for COD. Demonstrating remarkable efficacy, SINet surpasses numerous state-of-the-art object detection benchmarks, including ANet, by achieving an unprecedented accuracy of 91.18% from a baseline of 81.23%.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
Funder: Northern Border University
Award: G548247995
OpenAlex ID: W4411110581