ODExAI: A Comprehensive Object Detection Explainable AI Evaluation

There is no standard way to evaluate whether AI explanations for object detection are actually good, which makes it difficult for organizations to choose the right explanation method for their needs. This paper introduces ODExAI, the first comprehensive framework to systematically assess these explanation techniques.

All explanation methods involve trade-offs that organizations must understand before deployment: Region-based methods like D-CLOSE are highly faithful to how the model actually works (82% accuracy) but are very slow (71 seconds), while CAM-based methods like G-CAME are much faster (0.54 seconds) but less reliable at explaining the model's true behavior (55% accuracy).

Faster explanations don't necessarily mean better explanations: The paper found that the quickest explanation method (G-CAME at 0.54 seconds) achieved the best localization accuracy (96.13%) but had significantly lower faithfulness scores, meaning it may not accurately represent what the model is actually using to make decisions.

Organizations should evaluate explanation methods based on their specific use case: The research demonstrates that choosing an XAI method requires considering three factors—how accurately it pinpoints what the model is looking at, whether it truthfully represents the model's decision-making, and how fast it needs to run—rather than optimizing for just one dimension.

Evaluation benchmarks are now publicly available to help practitioners make informed decisions: The authors have released their implementation and evaluation framework so that organizations deploying object detection systems can systematically test which explanation method works best for their particular application.