STRinGS: Selective Text Refinement in Gaussian Splatting
a day ago
- #Computer Vision
- #Text Refinement
- #3D Reconstruction
- STRinGS is a text-aware, selective refinement framework for 3D Gaussian Splatting (3DGS) reconstruction.
- It treats text and non-text regions separately, refining text regions first and merging them later for full-scene optimization.
- STRinGS improves text readability by 63.6% over 3DGS at just 7K iterations.
- A new dataset, STRinGS-360, is introduced for evaluating text readability in 3D reconstruction.
- The method uses a two-phase pipeline: targeted densification of text Gaussians and full scene refinement.
- STRinGS-360 includes diverse text scenarios like curved, tightly packed, and occluded text.
- OCR Character Error Rate (CER) is used to measure text readability.
- The framework and dataset aim to enhance 3D scene understanding in text-rich environments.