[LS2N_ROMAS_IPI]Raw-Earth 3D Printing Dataset

Résumé fourni par la source

Raw-Earth 3D Printing Dataset The Raw-Earth 3D Printing Dataset is an open dataset developed from visual recordings acquired during an experimental campaign on robotic extrusion-based 3D printing of earthen materials. The original experiments were conducted by the SmartAMP research project through the work "Towards Predictive Models of Mechanical Performance in 3D-Printed Earthen Materials: The Role of Shrinkage and Cracking" to investigate the influence of material formulations and process parameters on the mechanical behavior and printability of 3D-printed earthen structures. Building upon these experimental recordings, we curated, organized, and standardized the visual data together with the associated experimental metadata to create an image dataset specifically designed for computer vision research, additive manufacturing, and real-time process monitoring of earth-based 3D printing. The experimental campaign investigated multiple earth formulations with different proportions of Bouyer Leroux raw earth, water, sand (0/1 mm and 0/0.2 mm), limestone filler, and flax fibers. Various printing parameters were also explored, including deposition speed, material flow rate (pump pressure), and material composition. Environmental conditions (temperature and relative humidity) and geometric characteristics of the printed specimens were recorded for every experiment. The experiments were performed using a UR10 robotic arm equipped with a screw-extrusion printing system (MAI Pictor pump). During printing, two Intel RealSense D405 RGB cameras continuously monitored the process from two synchronized viewpoints: Lateral view ("lateral"), observing the side profile of the deposited layers. Bottom/Top view ("dessus"), monitoring the deposition process from above the printing surface. Both cameras recorded images at a resolution of 1280 × 720 pixels and 5 fps, providing continuous visual monitoring throughout the printing process. Calibration grids installed on the printing table allow the images to be related to physical dimensions. The database contains recordings from 23 different material formulations, covering a broad range of printable earthen mixtures. Each experiment is identified by a formulation code (e.g., T100W40, TC45W23Sm55F0.75), together with its corresponding material composition, printing parameters, environmental conditions, shrinkage measurements, and acquisition video name. These metadata are provided in Base de données_exploitable.xlsx. Dataset Organization The dataset contains several subsets designed for different computer vision tasks. 1. Raw-earth patches: Raw-earth patches(579).zip This archive contains 579 grayscale texture patches extracted from the recorded printing images. Each patch has a fixed width of 200 pixels, while the height varies according to the extracted interlayer region. These patches represent raw earth textures captured during printing under different formulations and operating conditions. 2. Crossvalidation(579patchs_terre).zip This archive provides a 5-fold stratified cross-validation split for texture classification experiments. The dataset contains five texture categories with the following class distribution: Fluid(161 images,proportion 27.81%) Good(311 images,proportion 53.71%) Dry(31 images,proportion 5.35%) Tearing(56 images,proportion 9.67%) Geometric defect Écrasé (French) / Crushed (English) (20 images, proportion 3.45%) The predefined folds enable reproducible benchmarking of texture classification algorithms. 3. Interlayer segmentation dataset: Interlayer.zip This archive contains complete printing images together with manually annotated binary masks corresponding to the interlayer boundaries. The dataset includes images acquired from both camera viewpoints: dessus (top/bottom view) lateral (side view). For each image, a corresponding binary ground-truth mask identifies the interlayer regions. 4. Interlayer segmentation: crossvalidation_interlayer_terre.zip This archive provides predefined 5-fold cross-validation splits for supervised interlayer segmentation. The data are intended for training and evaluating deep learning models capable of automatically detecting interlayer boundaries during the printing process. 5. Texture semantic segmentation: crossvalidation_texturation_terre.zip This archive contains the semantic segmentation dataset for end-to-end texture recognition. The segmentation masks were generated by reconstructing full-size annotation maps from the manually labelled texture patches, enabling dense pixel-wise prediction of earth texture classes. The annotation masks contain the following classes and RGB color encoding: Fluid – Blue BGR (255, 0, 0): regions where the extruded material remains in a fluid state. Good – Green BGR (0, 255, 0): well-printed regions with satisfactory surface quality. Dry – Red BGR (0, 0, 255): regions where the material is excessively dry. Tearing – Cyan BGR (255, 255, 0): regions exhibiting tearing or discontinuities during deposition. Ecrase – Magenta BGR (255, 0, 255): regions where the deposited material has been excessively compressed or flattened. Background – Gray BGR (128, 128, 128): pixels outside the printed material. Interlayer – Black BGR (0, 0, 0): annotated interlayer boundaries separating successive deposited layers. The predefined 5-fold cross-validation splits ensure reproducible evaluation and comparison of semantic segmentation models. Metadata The file Base de données_exploitable.xlsx contains the experimental metadata associated with each printing trial, including: formulation identifier; earth, water, sand, limestone filler and flax fiber proportions; nozzle diameter; deposition speed; pump pressure; ambient temperature; relative humidity; specimen dimensions; linear shrinkage measurements; acquisition video names. These metadata facilitate correlations between material formulation, process parameters and image-based observations. Citation If you use this dataset in your research, please cite both this dataset and the associated publication describing the experimental campaign. License This dataset is distributed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0). https://creativecommons.org/licenses/by-nc-sa/4.0/