Voxel-Mesh Network for Geodesic-Aware 3D Semantic Segmentation of Indoor Scenes.

Area of Science:

• Computer Vision
• Artificial Intelligence
• 3D Deep Learning

Background:

• Sparse voxel-based methods using 3D Convolutional Neural Networks (CNNs) are state-of-the-art for 3D semantic segmentation.
• These methods struggle with spatially close objects and complex geometries due to a lack of geodesic information and geometric awareness.

Purpose of the Study:

• To introduce the Voxel-Mesh Network (VMNet), a novel 3D deep architecture.
• To leverage both voxel and mesh representations for improved 3D semantic segmentation of indoor scenes.

Main Methods:

• VMNet integrates Euclidean information from voxels (for contextual cues) and geodesic information from meshes (for separating disconnected surfaces).
• It employs an intra-domain attentive module for feature aggregation and an inter-domain attentive module for feature fusion.

Main Results:

• VMNet achieves superior performance on the ScanNet dataset, outperforming SparseConvNet and MinkowskiNet in mean Intersection over Union (mIoU).
• VMNet demonstrates higher accuracy (74.6% mIoU) compared to SparseConvNet (72.5%) and MinkowskiNet (73.6%).
• The network is more efficient, utilizing fewer parameters (17M) than SparseConvNet (30M) and MinkowskiNet (38M).

Conclusions:

• VMNet effectively combines voxel and mesh representations for enhanced 3D semantic segmentation.
• The proposed architecture offers a more efficient and accurate solution for segmenting complex indoor scenes.