IPNet: Polarization-based Camouflaged Object Detection via dual-flow network [dataset]
Description
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.
Files
Steps to reproduce
We have systematically constructed a comprehensive, authentic and challenging polarisation camouflage target detection dataset—PCOD_1200. This dataset comprises a total of 1,200 manually designed authentic camouflage scene samples. All data was captured using the Lucid Triton/HR-12000-S-PC polarisation camera in real-world conditions to ensure data authenticity and reliability. Within the defocused-plane polarisation camera system, raw image data is acquired in a defocused-plane format. In the PCOD_1200 dataset, these defocused plane images undergo specific algorithmic processing to yield raw polarised images at four orientations: 0°, 45°, 90°, and 135°. Thus, the defocused plane image essentially integrates these four images into a single high-resolution image according to the polarisation grating distribution pattern. This processing effectively generates an image comprising 1224 × 1024/2048 × 1500 macro-pixel units. Each macro-pixel composed of 2 × 2 basic pixels, thereby supplementing the original image data. RGB, DoLP, S1, S2, and other parameters are calculated via the Stokes vector. The code in the above folder pertains to polarimetric image processing, with detailed explanations provided in README.md.
Institutions
- Hefei University of Technology