Exploring Point Cloud Attention: Unraveling its Impact on Point Cloud Shape Completion
Principles of Deep Learning Course Project
Motivation
Existing attention visualizations are too local, showing only single query-key interactions. We need a global view of all query relationships to better understand how the model perceives overall shape and context.
Contribution & Method
Proposed a global attention visualization for point clouds using the PoinTr model as a baseline.
Analyzed the visualization, finding that the encoder consistently focuses on the edges of the point cloud.
- Encoder
- Decoder
Tested an edge-aware sampling method (MDS) based on this finding.
Results & Conclusion
The model with MDS sampling had lower quantitative scores, likely from over-sampling edges. However, it produced qualitatively better results with fewer outliers and cleaner shapes. This reveals an important trade-off between metric performance and the structural quality of the completed point cloud.
Despite the lower quantitative scores, we observed a significant qualitative improvement. The point clouds completed using MDS sampling showed a clear tendency to have fewer outliers and cleaner, more defined edges. This suggests that while edge-focused sampling may hurt general metrics, it helps in generating more structurally coherent shapes.
