Neural Randering Dataset Collection

Sammanfattning

Neural rendering methods such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting have transformed novel view synthesis, yet their development relies heavily on standardized datasets for training and evaluation. This thesis addresses the dataset gap by developing a multi-camera capture system and collecting two neural rendering datasets: a studio object dataset and a large-scale outdoor heritage site reconstruction. We designed and implemented an automated capture system using twelve synchronized industrial cameras arranged on a quarter-hemisphere arc with a motorized turntable. The system captures 432 images per object across 36 rotation positions. Custom C++ software controls camera synchronization, turntable automation, and data organization. The studio dataset contains 15 objects spanning diverse materials including diffuse, glossy, transparent, and reflective surfaces. For large-scale capture, we documented Gränsö Castle in Sweden using 5,262 images from drone and SLR photography. This dataset tests neural rendering scalability on architectural heritage sites with varying lighting conditions. We compared traditional photogrammetry (RealityCapture) against neural rendering methods (Nerfacto, Splatfacto) on both datasets. Results show that Splatfacto outperforms Nerfacto with average PSNR of 33.27 dB versus 22.07 dB on studio objects. Neural methods successfully reconstruct challenging materials (glass, metal) where photogrammetry fails due to feature matching limitations on specular surfaces. For large-scale scenes, photogrammetry remains practical for complete geometric models while neural rendering excels in bounded high-quality visualization. Both datasets and the capture software are released as open-source resources for the research community.