Master's Thesis | Linköping University | 2025
Author: Shaoxuan Yin
Supervisor: Sergey Ignatenko
Examiner: Jonas Unger
Programme: Computer Science, Master's Programme (6MICS)
Course: TQDV30 - Degree Project (30 hp)
This thesis investigates multi-view capture systems for neural rendering, comparing traditional photogrammetry with modern neural rendering methods (NeRF and 3D Gaussian Splatting). Two datasets were created: a controlled studio dataset of 15 objects with 432 images each, and a large-scale outdoor dataset of Gränsö Castle with 5,262 images. The work addresses how neural rendering handles view-dependent effects (reflections, transparency) and examines scalability limits for large scenes.
| Resource | Link |
|---|---|
| Thesis PDF | Download |
| Defense Slides | View Online |
| Capture Software | CamMatrixCapture |
| Studio Objects Dataset | Internet Archive |
- 15 objects captured with a 12-camera synchronized rig
- 432 images per object (12 cameras × 36 turntable positions)
- Includes challenging materials: glass, metal, fur, translucent objects
- ArUco marker-based geometric alignment
- Large-scale outdoor heritage site
- 5,262 images (drone + SLR photography)
- Multi-scale reconstruction from aerial to ground-level detail
The CamMatrixCapture software was developed for this thesis to control the multi-camera capture system. Features include:
- Synchronized capture from 12 Teledyne FLIR cameras
- Parallel image transfer (1.85× speedup)
- Automated turntable control
- Real-time preview and camera configuration
@mastersthesis{yin2025neural,
author = {Yin, Shaoxuan},
title = {Neural Rendering Dataset Collection},
school = {Linköping University},
year = {2025},
type = {Master's thesis},
number = {LiU-ITN-TEK-A--25/070--SE},
address = {Norrköping, Sweden}
}The thesis document is available for academic and educational purposes. The dataset is released under CC BY 4.0.