IPNet: Polarization-based Camouflaged Object Detection via dual-flow network [dataset]

In order to establish a universal standard for Polarization-based camouflaged object detection, we have meticulously constructed a comprehensive, real-world, and challenging dataset, dubbed PCOD_1200. The PCOD_1200 dataset consists of a total of 1200 instances of camouflage object detection scenarios, with 970 scenes allocated for training and the remaining 230 for testing. Different from widely utilized COD datasets, which were collected from the Internet through the Google search engine using specific keywords, our dataset is meticulously curated from authentic real-world images captured using a polarization camera within carefully designed camouflage scenarios. We draw inspiration from a wide array of authentic camouflage scenarios found in nature, military contexts, industrial environments, and daily life. Utilizing elements such as plants, specimens, Ghillie suits, and simulation models, we meticulously design and construct camouflage object detection scenes across diverse backgrounds including grasslands, sandy terrains, snowy landscapes, barren woods, and camouflaged fabrics. Specifically, it encompasses a broad assortment of scenes categorized into 8 main classes, further subdivided into 89 subclasses. Additionally, PCOD_1200 encompasses a considerable number of outdoor environments, including grasslands, dead leaves, dead trees, and other habitats, to further consider practical applications. Moreover, the camouflage scenes within PCOD_1200 encapsulate a variety of environmental lighting conditions, captured through passive detection mechanisms that forego any introduction of supplementary polarized light sources. This methodological choice is driven by the intent to authentically replicate natural conditions and real-world applications. To ensure the accurate representation of scene colors, each imaging scenario within the PCOD_1200 dataset is calibrated for white balance using Datacolor’s 24-color standard color card.

为建立偏振伪装目标检测（Polarization-based camouflaged object detection）的通用标准，我们精心构建了一款综合性、兼具真实场景与挑战性的数据集，命名为PCOD_1200。该数据集共包含1200组伪装目标检测场景样本，其中970组用于训练集，剩余230组用于测试集。与当前广泛应用的伪装目标检测（Camouflaged Object Detection, COD）数据集不同，现有COD数据集多通过谷歌搜索引擎使用特定关键词从互联网采集获取，本数据集的所有样本均为在精心设计的伪装场景中，使用偏振相机实拍得到的真实图像，经严格筛选整理而成。我们的场景设计灵感源自自然界、军事场景、工业环境与日常生活中各类真实伪装场景，通过植物、标本、吉利服（Ghillie suits）及仿真模型等元素，在草原、沙地、雪地、贫瘠林地、伪装织物等多样背景下，精心设计搭建了伪装目标检测场景。具体而言，该数据集涵盖8个主类别，进一步细分为89个子类别的丰富场景。此外，为贴合实际应用需求，PCOD_1200还包含大量户外生境场景，如草地、枯叶、枯木等。值得注意的是，PCOD_1200中的伪装场景覆盖了多种环境光照条件，且全部采用被动探测机制采集，未引入任何额外偏振光源，这一采集方案旨在真实还原自然环境与实际应用的原始状态。为确保场景色彩的准确还原，PCOD_1200数据集的每一组成像场景均使用Datacolor公司的24色标准色卡完成白平衡校准。