CDATP Fabric Database

Open PBR Textile Surface Database Dataset Description The Open PBR Textile Surface Database is an open, high-resolution image dataset for the digital representation and analysis of textile surfaces. It comprises 406 different textile materials, captured at 3840 × 3840 pixels at 605 dpi. For each material, both the front and back surfaces are available and identified as S1 (front) and S2 (back). The data were captured using a Vizoo xTex system and processed as Physically Based Rendering (PBR) maps. Up to six maps are available for each material surface: Alpha Base Color Displacement Roughness Metalness Normal PBR-based surface representation A central characteristic of the dataset is the representation of each textile surface through multiple PBR maps rather than conventional RGB imagery alone. The individual maps describe different aspects of the same surface and therefore provide complementary information. The Base Color map primarily represents visible color and pattern information. In contrast, maps such as Roughness, Normal, and Displacement represent surface characteristics that may not be visible, or may only be partially visible, in a conventional color image. For example, a printed pattern can be clearly visible in the Base Color map while producing little or no corresponding information in the Roughness map. Conversely, differences in surface structure, micro-relief, or reflectance behavior may become apparent in other PBR representations even when they are difficult to identify in the color image. The PBR maps therefore provide more than alternative visualizations of the same image. They represent different information channels of the same physical surface, allowing textile characteristics to be investigated from complementary perspectives. For rendering applications, the maps can be combined to reconstruct the appearance of a textile material under different lighting and viewing conditions. For research and machine-learning applications, they can be treated as separate channels or as a multi-modal representation of the same material. This enables applications such as material recognition, surface classification, texture analysis, surface reconstruction, texture synthesis, and the investigation of relationships between visible patterns and surface properties. The PBR maps were primarily generated for rendering applications. Their availability as high-resolution, structured image data additionally enables their use in computer vision and machine-learning workflows. Previous research has demonstrated the potential of image-based methods for the analysis of textile surface conditions and damage in the context of textile recycling [1]. The present dataset does not contain damaged textile products and is not intended as a damage-detection dataset. Rather, its multi-map representation provides a general-purpose basis for future research into textile surface analysis and other image-based applications. High-resolution source data and derived patches The original 3840 × 3840 pixel resolution allows the material surfaces to be subdivided into image patches of different sizes. For example, the maps can be divided into 64 × 64 or 128 × 128 pixel patches, depending on the requirements of a particular application. For a single map type, this provides a potential of approximately 60,000 individual image patches. Across all six map types, the theoretical potential is approximately 360,000 patches. These patches are derived from the original high-resolution material data and do not represent additional independent materials. The availability of the complete high-resolution surfaces allows researchers to generate task-specific patch datasets while retaining the original material-level information. Semantic annotations Additional semantic annotations are available for a subset of the materials. These annotations were created using Label Studio and describe characteristics including pattern, texture, and degree of wrinkling. For the respective materials, additional views or annotations from a 45° perspective are included. The detailed annotation scheme, label definitions, and annotation structure are provided in a separate documentation file accompanying the dataset. Scope The database deliberately follows an application-independent approach. The fiber composition and specific material composition of the textiles were not systematically investigated and are therefore not part of the dataset. The focus is on the digital representation of textile surfaces and their visible and PBR-relevant characteristics. The dataset is not restricted to a particular textile product category, production process, or research task. Potential applications include computer vision, machine learning, material recognition, surface analysis, texture synthesis, material reconstruction, digital material libraries, rendering, virtual product development, and the development and evaluation of AI-based methods. Data structure Each material is assigned a unique material identifier. The two surfaces are distinguished as S1 and S2 and 45 for the 45° pictures. The available PBR maps are uniquely associated with the corresponding material and surface. The dataset is accompanied by documentation describing the file structure, PBR map types, annotation scheme, and relevant metadata. Aim and contribution The aim of the database is to provide an open, high-resolution, reusable dataset of textile surfaces that is not restricted to a single application or industrial use case. Existing textile image collections often focus on patterns, prints, product representations, or production inspection. The present dataset instead combines high-resolution textile surface representations, front and back surfaces, multiple PBR channels, and semantic annotations. This combination provides access not only to information about what a textile looks like, but also to complementary representations of its surface characteristics. The dataset is intended as a foundation for the development and evaluation of future methods for the digital analysis, classification, reconstruction, synthesis, and generation of textile surfaces. Reference [1] F. Kunzelmann and Y. Kyosev, “Evaluation of Image-Based Methods for Damage Detection in the Recycling of Textile Products,” in Advances in the Textile and Clothing Research, Y. Kyosev, C. Cherif, O. Kyzymchuk, K. Hesse, and P. Penzel, Eds. Cham: Springer Nature Switzerland, 2026, pp. 283–300. doi: 10.1007/978-3-032-26592-0_19 .