SGPN paper explained

Similarity Group Proposal Network (SGPN) for 3D Point Cloud Instance Segmentation

Instance Segmentation results on point clouds using SGPN. Different colours represent different instances. (a) Instance segmentation on complete real scenes. (b) Single object part instance segmentation. (c) Instance segmentation on point clouds obtained from partial scans

What is 3D Instance Segmentation?

Motivation behind SGPN

Main Contribution of the paper

Model Architecture

The pipeline of the system for point cloud instance segmentation
Total loss is the sum of the Similarity loss, Confidence loss, and Semantic Segmentation loss

Branch 1: Similarity learning branch

Branch 2: Confidence learning branch

Branch 3: Semantic Segmentation branch

GroupMerging Algorithm for pruning proposals:

Results

Stanford 3D Indoor Semantic Segmentation Dataset (3DIS)

ShapeNet part segmentation

Limitations

Conclusion

Data Scientist/Software Engineer | Here to share some stories of all things tech | linkedin.com/in/zarifaziz/