Core Concepts
A graph neural network (SG-PGM) is proposed to solve the partial graph matching problem for 3D scene graph alignment, by fusing semantic and geometric features and enabling explicit partial matching.
Abstract
The paper presents SG-PGM, a graph neural network for 3D scene graph alignment. Key highlights:
- Defines 3D scene graph alignment as a partial graph matching problem and solves it with a graph neural network.
- Proposes a Point to Scene Graph Fusion (P2SG) module to combine semantic and geometric features for node embedding.
- Employs a soft top-k method to enable explicit partial matching, improving alignment accuracy.
- Introduces a Superpoint Matching Rescoring method that uses the scene graph alignment to guide point cloud registration, reducing false correspondences.
- Revisits the strategies for leveraging scene graph alignment in downstream tasks like overlap checking, point cloud registration, and mosaicking.
- Experiments show SG-PGM outperforms the previous state-of-the-art method SGAligner, especially in low-overlap and dynamic scenes.
Stats
The 3RScan dataset is used for evaluation, with 15,277 training and 1,882 validation samples.
Metrics like Chamfer Distance, Relative Rotation/Translation Error, Feature Matching Recall, and Registration Recall are reported for point cloud registration.
Quotes
"We reuse the geometric features learned by a point cloud registration method and associate the clustered point-level geometric features with the node-level semantic feature via our designed feature fusion module."
"We further propose a point-matching rescoring method, that uses the node-wise alignment of the 3D scene graph to reweight the matching candidates from a pre-trained point cloud registration method."