Segmentation and Smoothing Affect Explanation Quality More Than the Choice of Perturbation-based XAI Method for Image Explanations

Segmentation and smoothing matter more than method choice: The study found that how images are divided into regions (segmentation) and how explanations are smoothed have far greater impact on explanation quality than which specific perturbation-based method is used, challenging the common focus on comparing different XAI methods.

Attribution calculation is less important than assumed: Contrary to typical research emphasis, the way explanations are calculated and attributed has minimal impact on results, suggesting the field may be investing effort in optimizing the wrong parameters.

Per-pixel versus per-segment decisions are critical: Whether explanations are generated at the pixel level or segment level significantly affects quality, yet this parameter is rarely examined in existing research on image explanation methods.

Technical implementation details drive real-world performance: The findings suggest that practitioners implementing explanation systems should prioritize careful choices about segmentation algorithms and smoothing techniques over selecting between popular XAI methods, as these "boring" implementation details have outsized effects on explanation reliability.

Black-box model explanations require careful parameter tuning: Since perturbation-based methods work on any model without needing internal access, their widespread use for explaining opaque AI systems makes it especially important to understand which configuration choices actually improve explanation quality.