Camouflaged Image Synthesis Is All You Need to Boost Camouflaged Detection

TL;DR

The paper tackles data scarcity in camouflaged object detection by introducing SCODE, a GAN-based data synthesis framework that disentangles foreground camouflage from background. It employs a camouflage environment generator and a camouflage distribution classifier to synthesize realistic camouflaged images and corresponding masks, supervised by a camouflage distribution loss $\mathcal{L}_{cam}$. Through extensive experiments on CAMO, CHAMELEON, and COD10K, SCODE achieves state-of-the-art improvements and demonstrates strong ablations supporting the importance of the noise-adding module and camouflage classifier. The work also demonstrates plug-and-play data augmentation capabilities and analyzes limitations when generalizing to unseen non-camouflage domains, highlighting practical impact for improving camouflaged detection in real-world scenarios.

Abstract

Camouflaged objects that blend into natural scenes pose significant challenges for deep-learning models to detect and synthesize. While camouflaged object detection is a crucial task in computer vision with diverse real-world applications, this research topic has been constrained by limited data availability. We propose a framework for synthesizing camouflage data to enhance the detection of camouflaged objects in natural scenes. Our approach employs a generative model to produce realistic camouflage images, which can be used to train existing object detection models. Specifically, we use a camouflage environment generator supervised by a camouflage distribution classifier to synthesize the camouflage images, which are then fed into our generator to expand the dataset. Our framework outperforms the current state-of-the-art method on three datasets (COD10k, CAMO, and CHAMELEON), demonstrating its effectiveness in improving camouflaged object detection. This approach can serve as a plug-and-play data generation and augmentation module for existing camouflaged object detection tasks and provides a novel way to introduce more diversity and distributions into current camouflage datasets.

Figures (10)

• Figure 1: The figure shows a set of synthesized camouflage images generated by our Synthetic CamOuflage DEtection (SCODE) model to address the scarcity of data quantity and quality in camouflaged object detection tasks. The gray box represents the input object used by SCODE, the orange box contains generated camouflaged images, and the blue box shows the predicted masks generated by our camouflaged object detection models trained on the synthetic dataset. The synthetic dataset generated by SCODE improves the performance of camouflaged object detection models.
• Figure 2: Qualitative comparison of our model with state-of-the-art generative models. The four examples shown illustrate various scenarios of camouflage in nature, and our model successfully produces plausible results that blend in with the backgrounds. In contrast, the synthesized rabbit and fish in the bottom row by stable diffusing model rombach2021highresolution do not blend in, while the top row has no foreground objects at all. The results of StyleGAN2 Karras2019stylegan2 exhibit artifacts and are chaotic. Please zoom in to see the details.
• Figure 3: The architecture of our method. The upper part shows our camouflage environment generator, which takes an input foreground object and synthesizes a camouflaged environment to help the object blend in with its background. The left-down part illustrates the training process of the Camouflage Classifier, and the right-down part shows the dataset synthesis process. Note that "CAM" denotes camouflage and "NORM" denotes normal images.
• Figure 4: The Results of Generative Diversity Experiment. SCODE exhibits diverse results on both trained and untrained samples. Because data augmentation should be applied on all images in the training set of detection models, we visualize diversity results on the training set in the first two rows, and the testing set in the last two rows.
• Figure 5: Background Variation on Test Set. Backgrounds generated by our method exhibit variations compared to ground truth.