平滑的
计算机科学
人工智能
计算机视觉
代表(政治)
图像(数学)
政治学
法学
政治
作者
Donghyun Lee,Jinhoon Wang,Beomsu Cho,Junghyun Oh
出处
期刊:
日期:2024-10-29
卷期号:: 1556-1559
标识
DOI:10.23919/iccas63016.2024.10773027
摘要
The traditional dense visual SLAM system, which aims to generate precise and detailed maps, targets realtime reconstruction of large-scale indoor scenes. Contradictively, major recent dense SLAM methods utilizing neural implicit representations show promise in representing unobserved 3D spaces, but encounter difficulties in accuracy due to depth values from RGB-D sensors. Despite various advancements, depth images generated still suffer from noise, particularly in object edges and low-texture areas, impacting depth accuracy and camera pose tracking. To address this issue, this study proposes a new loss function, inspired by depth normalization in DN-Splatter, to suppress abrupt depth changes and enhance overall rendering performance, enabling more accurate tracking. By computing gradients of RGB and depth images and performing adaptive depth smoothing, our method effectively improves tracking performance, as demonstrated through comparisons using TUM RGB-D. It enhances sharpness of object contours and smoothness of surfaces, achieving an average improvement of 43% in root mean square error (RMSE) of trajectory evaluation, and contributing to the advancement of neural implicit representation-based SLAM.
科研通智能强力驱动
Strongly Powered by AbleSci AI