동국대학교 개별연구 - CSC 게임엔진을 활용한 딥러닝 분석용 데이터 생성 연구
Dongguk Univ CSE Independent Capstone Design - Research on Generating Deep Learning Analysis Data Using the Game Engine
SyntheticMultiView is a dataset generation pipeline built with Unreal Engine 5.5, designed to capture photometric, multi-view images of posed 3D scenes, along with pixel-level segmentation masks.
This project aims to generate high-quality synthetic image data for tasks such as 3D reconstruction, segmentation, and view-consistent perception. By using a rigged hand-object setup in Unreal Engine, the system captures the same scene from multiple predefined camera viewpoints, exporting both:
- RGB images (photorealistic)
- Segmentation masks (with transparent background)
Each frame is rendered with consistent lighting, camera configuration, and visual fidelity, making it suitable for multi-view learning or synthetic supervision.
- Multi-view rendering from fixed camera positions
- Pixel-wise segmentation masks (hand/object/background separation)
- Photometric lighting and realistic materials
- Fully reproducible within UE5.5 using packaged blueprint setup
- Unreal Engine 5.5
- Blender 4.22 (used for asset preprocessing and rigged animations)
- RenderTarget2D and custom post-process material setup for mask generation
- Adding joint (pose) or Landmark coordinate export
- Occlusion-aware visibility annotation
- Extension to topology-aware reconstruction or NeRF-compatible data