核心概念
Neural implicit representation enhances semantic mapping accuracy and real-time tracking in SNI-SLAM.
要約
SNI-SLAM is a semantic SLAM system that leverages neural implicit representation to improve dense visual mapping and tracking accuracy while providing semantic mapping of the entire scene. The system introduces hierarchical semantic representation for multi-level comprehension, integrating appearance, geometry, and semantic features through cross-attention for enhanced feature collaboration. By utilizing an internal fusion-based decoder, accurate decoding of semantic, RGB, and TSDF values is achieved from multi-level features. Extensive evaluations on Replica and ScanNet datasets demonstrate superior performance over existing NeRF-based SLAM methods in terms of mapping, tracking accuracy, and semantic segmentation.
統計
NeRF: Neural Radiance Fields (NeRF)
TSDF: Truncated Signed Distance Field (TSDF)
Replica dataset: 8 synthetic scenes with ground truth annotations
ScanNet dataset: 4 real-world scenes with ground truth annotations
ATE Mean[cm]: Average Trajectory Error in centimeters
mIoU(%): Mean Intersection over Union percentage
引用
"Our SNI-SLAM leverages the correlation of multi-modal features in the environment to conduct semantic SLAM based on Neural Radiance Fields (NeRF)."
"We propose a feature collaboration method between appearance, geometry, and semantics."
"Our SNI-SLAM method demonstrates superior performance over all recent NeRF-based SLAM methods."