An adaptable methodology for generating realistic annotated synthetic data with 3D Gaussian splatting and modeling software

The increasing demand for large annotated datasets in computer vision underscores the need for scalable synthetic data generation methods, as traditional approaches often lack adaptability or offer limited annotations. To address these challenges, this work introduces a comprehensive methodology that leverages 3D Gaussian splatting and Blender to generate photorealistic synthetic datasets with automatic annotations, suitable for training deep learning computer vision models in multiple tasks, including object detection, pose estimation, semantic and instance segmentation, depth estimation, and surface normal estimation. Designed for broad applicability, this approach was developed with the significant challenges of industrial scenarios in mind, where data scarcity impedes the development of resilient AI-based situational awareness systems. The proposed pipeline can model any target scenario by reconstructing high-fidelity 3D environments, integrating synthetic objects through diverse modeling techniques, and applying procedural generation to vary lighting, camera angles, and object placements. With all annotations generated automatically, this approach facilitates the efficient creation of extensive annotated datasets, thereby eliminating manual labeling efforts. To validate the effectiveness of our synthetic data, we train computer vision models exclusively on these datasets and evaluate their performance in real applications, demonstrating their generalization capability and practical relevance. The results confirm that our methodology provides a scalable, flexible, and high-quality alternative to traditional data collection, bridging the gap between synthetic and real-world datasets while offering a powerful tool for both research and industrial applications. Our code is publicly accessible at https://github.com/juanjesus-ldo/scene_generator .