XAI for Point Cloud Data using Perturbations based on Meaningful Segmentation

Explainability for 3D AI is increasingly important: The paper addresses a gap in understanding how AI systems make decisions about 3D objects (like cars, guitars, and laptops), which is critical as these technologies are deployed in real-world applications where transparency matters.

Meaningful explanations require human-interpretable segments: Rather than using arbitrary data perturbations, the method generates explanations based on meaningful parts of 3D objects that humans can actually understand, making it easier for non-experts to analyze why an AI system classified something a particular way.

A novel "point-shifting" technique improves explanation reliability: The paper introduces a new technical approach that removes points from 3D data in a way that ensures they truly stop influencing the AI's decision, addressing a limitation in previous explanation methods that may have produced misleading results.

Segmentation-based explanations bridge the gap between AI and human understanding: By using existing segmentation models to break down 3D objects into recognizable parts, the method creates explanations that align with how humans naturally think about objects, rather than producing abstract or difficult-to-interpret outputs.

This work supports responsible AI deployment in critical areas: By making AI decision-making processes more transparent and understandable for 3D object classification, the research enables better human oversight and more informed decision-making when these systems are used in important applications.