Unified resolution of target-conditioned fusion challenges in reference-guided COD

Develop a unified reference-guided camouflaged object detection framework that jointly performs target-conditioned foundation-feature selection, depth-matched detail reconstruction, and reliability-aware multi-reference aggregation.

Background

Reference-guided camouflaged object detection must identify target instances specified by visual references while suppressing semantically unrelated camouflaged objects and recovering weak boundaries and fine structures. The paper identifies three complementary requirements: selecting target-relevant evidence from frozen foundation-model features, injecting reference information at semantic depths appropriate to query features, and estimating the reliability of heterogeneous reference samples.

Although the proposed framework addresses these requirements through Reference-Conditioned Dual-Backbone Fusion, hierarchical reference fusion and decoding, and Cross-Reference Consensus Aggregation, the introduction explicitly characterizes their joint resolution within one framework as unresolved. This open problem concerns integrating all three capabilities coherently rather than treating them as isolated design components.