Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DGA-Net: Enhancing SAM with Depth Prompting and Graph-Anchor Guidance for Camouflaged Object Detection

Yuetong Li, Qing Zhang, Yilin Zhao, Gongyang Li, Zeming Liu

TL;DR

DGA-Net addresses the challenge of applying SAM to camouflaged object detection by introducing depth-guided prompting and a graph-based cross-modal fusion strategy. The Cross-modal Graph Enhancement (CGE) creates a unified RGB-depth representation by dynamically aligning texture-rich RGB features with depth-based structural cues via a heterogeneous graph attention mechanism, while the Anchor-Guided Refinement (AGR) injects a global semantic anchor and propagates it directly to all shallow layers to prevent information decay. The combination yields improved segmentation accuracy across standard COD benchmarks, outperforming state-of-the-art methods, including other SAM-based and depth-augmented approaches. The approach demonstrates that treating depth as a dense geometric prompt and enforcing cross-level top-down guidance can robustly guide SAM for fine-grained camouflage segmentation, with potential for broader open-world and industrial applications.

Abstract

To fully exploit depth cues in Camouflaged Object Detection (COD), we present DGA-Net, a specialized framework that adapts the Segment Anything Model (SAM) via a novel ``depth prompting" paradigm. Distinguished from existing approaches that primarily rely on sparse prompts (e.g., points or boxes), our method introduces a holistic mechanism for constructing and propagating dense depth prompts. Specifically, we propose a Cross-modal Graph Enhancement (CGE) module that synthesizes RGB semantics and depth geometric within a heterogeneous graph to form a unified guidance signal. Furthermore, we design an Anchor-Guided Refinement (AGR) module. To counteract the inherent information decay in feature hierarchies, AGR forges a global anchor and establishes direct non-local pathways to broadcast this guidance from deep to shallow layers, ensuring precise and consistent segmentation. Quantitative and qualitative experimental results demonstrate that our proposed DGA-Net outperforms the state-of-the-art COD methods.

DGA-Net: Enhancing SAM with Depth Prompting and Graph-Anchor Guidance for Camouflaged Object Detection

TL;DR

DGA-Net addresses the challenge of applying SAM to camouflaged object detection by introducing depth-guided prompting and a graph-based cross-modal fusion strategy. The Cross-modal Graph Enhancement (CGE) creates a unified RGB-depth representation by dynamically aligning texture-rich RGB features with depth-based structural cues via a heterogeneous graph attention mechanism, while the Anchor-Guided Refinement (AGR) injects a global semantic anchor and propagates it directly to all shallow layers to prevent information decay. The combination yields improved segmentation accuracy across standard COD benchmarks, outperforming state-of-the-art methods, including other SAM-based and depth-augmented approaches. The approach demonstrates that treating depth as a dense geometric prompt and enforcing cross-level top-down guidance can robustly guide SAM for fine-grained camouflage segmentation, with potential for broader open-world and industrial applications.

Abstract

To fully exploit depth cues in Camouflaged Object Detection (COD), we present DGA-Net, a specialized framework that adapts the Segment Anything Model (SAM) via a novel ``depth prompting" paradigm. Distinguished from existing approaches that primarily rely on sparse prompts (e.g., points or boxes), our method introduces a holistic mechanism for constructing and propagating dense depth prompts. Specifically, we propose a Cross-modal Graph Enhancement (CGE) module that synthesizes RGB semantics and depth geometric within a heterogeneous graph to form a unified guidance signal. Furthermore, we design an Anchor-Guided Refinement (AGR) module. To counteract the inherent information decay in feature hierarchies, AGR forges a global anchor and establishes direct non-local pathways to broadcast this guidance from deep to shallow layers, ensuring precise and consistent segmentation. Quantitative and qualitative experimental results demonstrate that our proposed DGA-Net outperforms the state-of-the-art COD methods.
Paper Structure (26 sections, 13 equations, 9 figures, 6 tables)

This paper contains 26 sections, 13 equations, 9 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Camouflaged Object Detection
  4. Segment Anything Model
  5. Graph-based Modeling in Camouflaged Object Detection
  6. Method
  7. Overall
  8. Cross-modal Graph Enhancement
  9. Graph Pooling
  10. Heterogeneous Graph Attention
  11. Anchor-Guided Refinement Module
  12. Semantic-Structural Anchor Generation
  13. Cross-Level Semantic Propagation
  14. Loss Function
  15. Experiments
  16. ...and 11 more sections

Figures (9)

  • Figure 1: Visualization of different prompting strategies in SAM-based camouflaged object detection. (a) SAM without any prompt. SAM2-UNet (b) Segmentation guided by a box prompt. SAM-Adapter (c) Dual-branch RGB-Depth encoder guided by box or hybrid box-depth prompts. DSAMSAM-DSA (d) Segmentation guided by a depth-aware geometric prompt and graph-based RGB–Depth fusion (Ours).
  • Figure 2: Overview of our proposed DGA-Net, which consists of four main components, i.e., SAM encoder, unified cross-modal encoder, anchor-guided refinement (AGR) and SAM decoder.
  • Figure 3: The details of our cross-modal graph enhancement (CGE) module.
  • Figure 4: Visual comparisons of some recent COD methods and our proposed network.
  • Figure 5: Visualization results of the ablation study on different key components.
  • ...and 4 more figures