研究目的
Investigating the effectiveness of a deep neural network for 3D semantic segmentation of raw point clouds by introducing multi-scale feature learning and global and local feature aggregation blocks.
研究成果
The 3DMAX-Net outperforms state-of-the-art methods in 3D point cloud semantic segmentation by effectively utilizing multi-scale spatial contextual features and local and global feature aggregation. It achieves a mean IoU of 47.5% and an overall accuracy of 79.5% on the S3DIS dataset.
研究不足
The computational complexity of processing irregular data structures with DSU and USU may slow down the processing. The network's performance on objects with similar planes (e.g., tables and floors) is limited.
1：Experimental Design and Method Selection:
The proposed 3DMAX-Net integrates multi-scale feature learning and local and global feature aggregation blocks for semantic segmentation of 3D point clouds.
2：Sample Selection and Data Sources:
The Stanford large-scale 3D Indoor Spaces Dataset (S3DIS) is used, containing 3D scans from 6 different large-scale indoor scenes with 13 semantic labels.
3：List of Experimental Equipment and Materials:
A computer with 48GB RAM, a GTX 1080 Ti GPU, and an Intel i7 6700K CPU was used for training and testing.
4：Experimental Procedures and Operational Workflow:
The scene data was split into blocks, with data augmentation applied. The network was trained using the Adam optimizer with early stopping.
5：Data Analysis Methods:
Performance was evaluated using Intersection over Union (IoU), average accuracy per class, and overall accuracy.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
