Uncertainty-Aware Transformer for Referring Camouflaged Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Referring camouflaged object detection (Ref-COD) is challenging due to similar backgrounds and feature discrepancies between targets and references.
• Existing methods struggle to effectively discern camouflaged objects and integrate visual references.

Purpose of the Study:

• To propose a novel uncertainty-aware transformer (UAT) for improved Ref-COD performance.
• To address the challenges of feature similarity and reference integration in Ref-COD.

Main Methods:

• UAT employs a cross-attention mechanism for aligning visual references with camouflaged features.
• A referring feature aggregation (RFA) module integrates reference information for targeted feature learning.
• A transformer probabilistic decoder (TPD) models patch dependencies probabilistically to capture uncertainty-aware features.

Main Results:

• UAT demonstrates superior performance on the Ref-COD benchmark compared to state-of-the-art methods.
• The model shows competitive results on conventional camouflaged object detection (COD) datasets, indicating scalability.

Conclusions:

• The proposed UAT effectively tackles Ref-COD challenges by integrating uncertainty modeling and reference guidance.
• UAT offers a robust and scalable solution for camouflaged object detection tasks.