Zhiwei Xing, Xiaorui Zhu, and Yudong Wu
Semantic segmentation, lightweight network, super-pixel extraction, semantic mapping
In this paper, a novel real-time 3D dense semantic mapping system, SemanticSurfel, is proposed to integrate semantic segmentation results, poses, depth graphs into the constructed map, and to scale well in large-scale environments. First, a lightweight semantic segmentation network, HybridNet, is designed with efficient Hybrid Basic Blocks and Hybrid Dilated Blocks in the encoder and Attention Pyramid Module in the decoder to accurately and efficiently segment the input image at pixel level. Then, super-pixels extracted from semantic, depth, and intensity graphs are used to construct surfels to build the 3D dense semantic map according to the pose graph of a sparse SLAM system. Extensive experiments were carried out to evaluate the performance of HybridNet and SemanticSurfel. Experimental results demonstrate that HybridNet achieves a good balance between accuracy and hybrid efficiency, and the SemanticSurfel system achieves great accuracy and scales well in large-scale environments.
Important Links:
Go Back
