GBC: Gaussian-Based Colorization and Super-Resolution for 3D Reconstruction
· Paper · Optical Flow / Super-Resolution / Colorization / COLMAP / 3D Gaussian Splatting
Published at ACM SIGGRAPH VRCAI 2024.
GBC transforms low-resolution monochrome video into higher-resolution, colorized frames suitable for geometrically accurate 3D reconstruction.
The framework combines Bidirectional Optical Flow Super-Resolution (BOF-SR) with Temporal Optical Flow-based Colorization (TOF-CO). BOF-SR recovers spatial detail while using information from neighboring frames; TOF-CO reduces frame-to-frame color inconsistency. The restored sequence is then processed with COLMAP and 3D Gaussian splatting to reconstruct the scene.
We also created a dataset for low-light, low-quality historical footage and evaluated the method alongside public datasets. The experiments focus on temporal coherence, color accuracy, and 3D reconstruction fidelity.