EGNet: 3D Semantic Segmentation Using Point, Voxel, and Mesh Data
Author Information
Author(s): Li Qi, Song Yu, Jin Xiaoqian, Wu Yan, Zhang Hang, Zhao Di, Prieto Javier
Primary Institution: School of Computer Science and Technology, Changchun University of Science and Technology, Changchun, China
Hypothesis
Can integrating Euclidean and geodesic features improve semantic segmentation accuracy in indoor scenes?
Conclusion
The EGNet effectively integrates Euclidean and geodesic features, leading to improved semantic segmentation results.
Supporting Evidence
- The EGNet achieved a mean intersection over union (mIoU) of 73.3% on the ScanNet validation set.
- EGNet outperformed existing methods like DCM-Net and SparseConvNet in semantic segmentation tasks.
- The integration of Euclidean and geodesic features led to improved boundary segmentation accuracy.
Takeaway
This study created a new method to help robots understand indoor spaces better by combining different types of data, making it easier to tell objects apart.
Methodology
The study developed a network called EGNet that processes point, voxel, and mesh data through Euclidean and geodesic branches for semantic segmentation.
Limitations
The study does not address the computational complexity of the proposed method in large-scale scenarios.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website