TR-NeRF: Transient-Robust Neural Radiance Fields for Indoor Scenes

Neural Radiance Fields (NeRF) have become popular for generating high-quality scene images from new viewpoints using captured multi-view static scene images. However, most NeRF-based methods excel at synthesizing new views from stable scenes; when interference objects are present in indoor scenes, these methods tend to produce artifacts and floating objects in the rendered views. To address these issues, we propose the TR-NeRF method, which utilizes multi-scale dynamic masks to eliminate interference objects in the scene while integrating spatial consistency and edge gradient detection to avoid misjudgments, thereby preserving scene edge details to the greatest extent when removing transient interference objects. Experimental results show that our method improves metrics such as PSNR and SSIM over baseline methods, with PSNR improvements reaching up to 0.5dB. Finally, our proposed method requires almost no hyperparameters and can be easily integrated with modern NeRF frameworks.