Gaussians in the City: Enhancing 3D Scene Reconstruction under distractors with Text-guided Segmentation and Inpainting
Naoki Shitanda, Jun Rekimoto · 2024
Figure 1: Comparison of Rendered Images: (a) ground truth image in dataset (b) rendered image from vanilla 3DGS from a new camera pose (c) ours after initial optimization (d) ours after inpainting and second optimization.Red circles illustrate defects and artifacts around heavily masked regions are suppressed in our method with inpainting.