Unsupervised Camouflaged Object Segmentation as Domain Adaptation

Yi Zhang, Chengyi WuWorkshops @ ICCV 2023 · DOI

Abstract

Camouflaged object segmentation (COS) is a challenging computer vision task due to the inherent ambiguity of object boundaries designed to deceive. While supervised methods have made significant strides, they rely on extensive, high-quality manual annotations. This paper discusses UCOS-DA, a novel unsupervised framework that treats COS as a source-free domain adaptation problem. By adapting self-supervised features from a generic source model (DINO) to the specific domain of camouflage using a Foreground-Background-Contrastive Self-Adversarial (FBA) module, the method achieves state-of-the-art performance on multiple benchmarks with minimal training data.

The Challenge of Camouflage

Camouflage is a biological and tactical strategy designed to deceive visual systems by breaking the boundaries that standard computer vision models rely on. Unlike salient objects that stand out from their environment, camouflaged objects utilize background matching, transparency, and disruptive patterns to obscure their presence.

This creates a significant "domain gap" for standard models. Most unsupervised segmentation models are pre-trained on generic datasets like ImageNet, where objects are typically distinct from their backgrounds. When these models encounter a tiger in tall grass or a spider on bark, they often fail to distinguish the object from its surroundings, leading to incomplete or failed segmentations.

Figure 1: Real-world examples of camouflage, including natural (tiger, spider, fish) and artificial (body paint) scenarios that challenge standard visual boundaries.

Methodology: UCOS-DA

To bridge this gap without human labels, the authors propose UCOS-DA (Unsupervised Camouflaged Object Segmentation as Domain Adaptation). The framework treats a frozen, self-supervised Vision Transformer (ViT) pre-trained with DINO as a "source" model and adapts it to the "target" domain of camouflage.

Key Components:

  • Source Model: A frozen DINO-based ViT that provides a baseline for "objectness."
  • FBA Module: The core innovation. Since camouflage boundaries are inherently ambiguous, the FBA module uses hierarchical linear layers to specifically weight and refine boundary-specific local features. It employs an adversarial strategy to force the model to find the subtle differences that give away a hidden object.
  • Pseudo-Labeling: The training is guided by MaskCut, which generates initial coarse guesses of object locations to kickstart the adaptation process.
Figure 2: The three primary camouflage mechanisms—background matching, transparency, and disruptive patterning—that the UCOS-DA model must learn to navigate.

Experimental Results

The efficiency of UCOS-DA is a standout feature. While supervised models often require thousands of labeled masks, UCOS-DA was trained on a tiny, unlabeled set of just 600 images. Despite this, it outperformed previous unsupervised leaders across all major benchmarks:

  • Quantitative Performance: On the COD10K benchmark, UCOS-DA achieved an mIoU of 0.462, significantly outperforming the previous best unsupervised method, FOUND (0.428).
  • Qualitative Success: Visual results demonstrate that the model effectively captures the intricate silhouettes of insects and reptiles that blend almost perfectly into bark or leaves—cases where previous models often produced "blobby" or incomplete masks.

Conclusion

By reframing unsupervised camouflage segmentation as a domain adaptation task, this research provides a practical path forward for specialized computer vision tasks where data is scarce. The success of a simple linear target model suggests that the "knowledge" of camouflage is already latent within large self-supervised models; it simply requires the right adaptation mechanism to be unlocked. Future work may extend this approach to other challenging domains, such as medical imaging or low-light surveillance.

Summary updated

BibTeX
@inproceedings{zhang2023unsupervised,
  author = {Yi Zhang and Chengyi Wu},
  title = {{Unsupervised Camouflaged Object Segmentation as Domain Adaptation}},
  booktitle = {Workshops @ ICCV},
  year = {2023},
  doi = {10.1109/iccvw60793.2023.00468},
}
Figure