While metrics such as precision, recall, and related measures are useful for evaluating object detection models, they provide a limited perspective on model behavior. Even with carefully prepared training data and robust optimization, there is no guarantee regarding what features a model actually learns. In practice, a model may associate certain background elements, i.e., scene level objects, with the presence of target classes, resulting in unintended contextual dependencies. Conventional performance metrics, however, do not reveal this issue. To address this gap, this paper introduces a black box explainability approach that evaluates object detection models by quantifying the influence of scene level objects on class identification. By comparing Average Precision (AP) on test data with and without specific scene elements, the method highlights the extent to which those objects contribute to model performance. Three experiments are presented to demonstrate the method’s utility. The findings provide both quantitative and global explanations of model behavior, yielding a more complete picture of object detection performance.
Vonderhaar et al. (Wed,) studied this question.