Toward Realistic Camouflaged Object Detection: Benchmarks and Method

Zhimeng Xin, Tianxu Wu, Shiming Chen, Shuo Ye, Zijing Xie, Yixiong Zou, Xinge You, Yufei GuoarXiv 2025 · OpenReview

Abstract

Camouflaged object detection (COD) primarily relies on semantic or instance segmentation methods. While these methods have made significant advancements in identifying the contours of camouflaged objects, they may be inefficient or cost-effective for tasks that only require the specific location of the object. Object detection algorithms offer an optimized solution for Realistic Camouflaged Object Detection (RCOD) in such cases. However, detecting camouflaged objects remains a formidable challenge due to the high degree of similarity between the features of the objects and their backgrounds. Unlike segmentation methods that perform pixel-wise comparisons to differentiate between foreground and background, object detectors omit this analysis, further aggravating the challenge. To solve this problem, we propose a camouflage-aware feature refinement (CAFR) strategy. Specifically, in CAFR, we introduce the Adaptive Gradient Propagation (AGP) module that fine-tunes all feature extractor layers in large detection models to fully refine class-specific features from camouflaged contexts. We then design the Sparse Feature Refinement (SFR) module that optimizes the transformer-based feature extractor to focus primarily on capturing class-specific features in camouflaged scenarios. To facilitate the assessment of RCOD tasks, we manually annotate the labels required for detection on three existing segmentation COD datasets, creating a new benchmark for RCOD tasks.

Introduction: From Segmentation to Detection

Traditionally, camouflaged object detection (COD) has been treated as a segmentation problem, focusing on identifying the exact pixel-level contours of hidden targets. While precise, segmentation is computationally expensive and requires high-effort manual annotation. In many practical applications—such as search-and-rescue, wildlife monitoring, or military surveillance—identifying the specific location (bounding box) and category of an object is often sufficient and significantly more efficient.

Comparison between traditional COD (segmentation) and the proposed RCOD (detection), highlighting the challenge of similar patch pairs.

This paper introduces Realistic Camouflaged Object Detection (RCOD), shifting the focus toward bounding-box-based detection. However, standard object detectors struggle in these environments because the foreground and background share nearly identical textures, colors, and patterns. As shown in Figure 1, "extremely similar patch pairs" make it difficult for models to distinguish between a camouflaged animal and its environment without the pixel-wise analysis inherent in segmentation.

Methodology: Camouflage-Aware Feature Refinement (CAFR)

To address the limitations of standard detectors, we propose the Camouflage-Aware Feature Refinement (CAFR) strategy. This approach leverages the prior knowledge of large pre-trained models and enhances their ability to perceive subtle distinctions in camouflaged contexts through two primary modules: Adaptive Gradient Propagation (AGP) and Sparse Feature Refinement (SFR).

The CAFR framework, featuring Adaptive Gradient Propagation (AGP) and Sparse Feature Refinement (SFR) modules integrated with a Transformer-based backbone.

Adaptive Gradient Propagation (AGP)

Standard fine-tuning often freezes the model's backbone, training only the final layers. However, for camouflaged objects, crucial details like subtle edge variations are often lost in high-level abstractions. AGP allows for the fine-tuning of all layers in the feature extractor. To prevent overfitting to the relatively small camouflaged datasets, AGP employs a "restriction factor" that limits gradient updates, allowing the model to "re-learn" low-level distinctions without losing its generalizable pre-trained knowledge.

Sparse Feature Refinement (SFR)

In a typical bounding box for a camouflaged object, the actual class-specific features may only occupy a small fraction of the area. This "feature sparsity" causes Transformer-based models to become confused by background patches within the box. SFR addresses this by creating synthetic training samples where bounding boxes are cropped and tiled into grids (e.g., $2\times2$ or $4\times4$). This multi-scale tiling forces the model's self-attention mechanism to focus on dense, class-specific features across different contexts, effectively training the model to ignore distracting background similarities.

New Benchmarks for RCOD

Because existing COD datasets are designed for segmentation, we have manually annotated three new benchmarks for the RCOD task. These datasets provide the bounding boxes and class labels necessary for evaluating detection performance:

  • COD10K-D: Based on COD10K-v2, containing 68 categories across 10,000 images.
  • NC4K-D: A detection-oriented version of the NC4K dataset (4,121 images).
  • CAMO-D: A detection-oriented version of the CAMO dataset (1,241 images).

Key Results and Insights

The CAFR strategy was evaluated using state-of-the-art large models, including GLIP and Grounding DINO. Our results demonstrate that CAFR significantly boosts detection accuracy in camouflaged scenes:

  • Performance Gains: On the COD10K-D dataset, CAFR improved the mean Average Precision (mAP) of GLIP-T by 2.4% and Grounding DINO-T by 1.7%.
  • Efficiency: RCOD tasks were found to be significantly more efficient than segmentation-based COD in terms of FLOPs, supporting the case for detection in real-time deployment scenarios.
  • Robustness: Qualitative analysis shows that CAFR-enhanced models produce fewer false positives (redundant detections) in highly textured backgrounds where standard models often "see" objects in every patch of camouflage.

Conclusion

By shifting the paradigm from segmentation to detection and introducing the CAFR strategy, this work provides a more practical and efficient path for identifying camouflaged objects. The combination of Adaptive Gradient Propagation and Sparse Feature Refinement allows large models to bridge the "cognitive gap" in camouflaged scenarios, while our new benchmarks provide a foundation for future research in Realistic Camouflaged Object Detection.

Summary updated

Abstract

Camouflaged object detection (COD) primarily relies on semantic or instance segmentation methods. While these methods have made significant advancements in identifying the contours of camouflaged objects, they may be inefficient or cost-effective for tasks that only require the specific location of the object. Object detection algorithms offer an optimized solution for Realistic Camouflaged Object Detection (RCOD) in such cases. However, detecting camouflaged objects remains a formidable challenge due to the high degree of similarity between the features of the objects and their backgrounds. Unlike segmentation methods that perform pixel-wise comparisons to differentiate between foreground and background, object detectors omit this analysis, further aggravating the challenge. To solve this problem, we propose a camouflage-aware feature refinement (CAFR) strategy. Since camouflaged objects are not rare categories, CAFR fully utilizes a clear perception of the current object within the prior knowledge of large models to assist detectors in deeply understanding the distinctions between background and foreground. Specifically, in CAFR, we introduce the Adaptive Gradient Propagation (AGP) module that fine-tunes all feature extractor layers in large detection models to fully refine class-specific features from camouflaged contexts. We then design the Sparse Feature Refinement (SFR) module that optimizes the transformer-based feature extractor to focus primarily on capturing class-specific features in camouflaged scenarios. To facilitate the assessment of RCOD tasks, we manually annotate the labels required for detection on three existing segmentation COD datasets, creating a new benchmark for RCOD tasks. Code and datasets are available at: https://github.com/zhimengXin/RCOD.
BibTeX
@misc{xin2025toward,
  author = {Zhimeng Xin and Tianxu Wu and Shiming Chen and Shuo Ye and Zijing Xie and Yixiong Zou and Xinge You and Yufei Guo},
  title = {{Toward Realistic Camouflaged Object Detection: Benchmarks and Method}},
  year = {2025},
  eprint = {2501.07297},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.07297},
}
Figure