Reproduce “Rethinking Detecting Salient and Camouflaged Objects in Unconstrained Scenes” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhangjun Zhou, Yiping Li, Chunlin Zhong, Jianuo Huang, Jialun Pei, Hua Li, He Tang
Researchers from Huazhong University of Science and Technology, CUHK, and Hainan University introduce a framework to address misclassification between salient and camouflaged objects in unconstrained scenes. This framework comprises USC12K, the first dataset with comprehensive multi-attribute labels, USCNet, a model explicitly modeling attribute relationships, and the CSCS metric, with USCNet achieving state-of-the-art performance and significantly reducing confusion.
Key insights
- The observed cross-task misinterpretation by existing SOD and COD models stems from the restrictive nature of current datasets and the absence of explicit attribute relationship modeling in network architectures.
- A comprehensively annotated dataset (USC12K) that embraces all logical scenarios of object existence can significantly mitigate the misdetection issue between salient and camouflaged objects.
- Explicitly modeling attribute relationships through a dual prompt query mechanism (Inter-SPQ for universal features, Intra-SPQ for sample-specific context) within a foundational vision model like SAM enables robust disentanglement of opposing visual attributes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.