GranSSG: Correlating Volumetric Granularities for 3D Semantic Scene Graph Prediction

Area of Science:

• Computer Vision
• Artificial Intelligence
• 3D Scene Understanding

Background:

• Predicting 3D Semantic Scene Graphs (3DSSG) is crucial for structured scene representation.
• Current methods face challenges with instance granularity discrepancies, limiting perception of varied object sizes.

Purpose of the Study:

• Introduce GranSSG, a novel approach for 3DSSG prediction.
• Integrate volumetric granular awareness to handle diverse instance scales effectively.

Main Methods:

• Volumetric Pooling block aggregates features from multiple instance volumes for multi-granularity pattern enhancement.
• Granularity Transformer block dynamically focuses attention on instance features across network layers.
• Cross-Granularity Correlation Transformer block adaptively fuses features to improve instance pair relationship prediction.

Main Results:

• GranSSG significantly enhances 3DSSG prediction performance.
• Demonstrated superior results on a challenging 3DSSG benchmark.
• Established a new state-of-the-art in 3DSSG prediction.

Conclusions:

• GranSSG effectively addresses granularity discrepancies in 3DSSG prediction.
• The proposed method offers a more comprehensive understanding of scene instances and their relationships.
• GranSSG represents a significant advancement in the field of 3D scene understanding.