SGS-SLAM introduces a unique semantic feature loss to address the limitations of traditional depth and color losses in object optimization. By leveraging multi-channel optimization, it prevents erroneous reconstructions caused by cumulative errors. The system excels in producing accurate and high-fidelity global reconstruction by capturing dense photometric information through differentiable rendering. Compared to NeRF-based methods, SGS-SLAM demonstrates remarkable superiority in rendering speed, reconstruction quality, and segmentation accuracy. The integration of semantic features within the method significantly advances optimal scene interpretation and precise object-level geometry. Extensive experiments show that SGS-SLAM provides state-of-the-art tracking and mapping results while maintaining rapid rendering speeds.
다른 언어로
소스 콘텐츠 기반
arxiv.org
더 깊은 질문