Gaussians in the city: Enhancing 3D scene reconstruction under distractors with text-guided segmentation and inpainting

Author

Shitanda, Naoki and Rekimoto, Jun

Abstract

We propose a novel method to reconstruct 3D Gaussian scene from images that contain both static and dynamic distractors, e.g., vehicles and pedestrians, by using text-guided segmentation and inpainting techniques. Our approach generates masks for distractors, detects heavily masked regions and inpaints them to suppress defects and artifacts for scene optimization while ignoring distractors.