On Evaluation of Vision Datasets and Models using Human Competency Frameworks
This addresses the need for more nuanced evaluation methods in computer vision research, though it is incremental as it adapts an existing framework from psychometrics.
The paper tackles the problem of coarse evaluation in computer vision by applying Item Response Theory (IRT) to infer interpretable latent parameters for models and dataset items, enabling richer analysis beyond accuracy, such as assessing model calibration and selecting informative data subsets.
Evaluating models and datasets in computer vision remains a challenging task, with most leaderboards relying solely on accuracy. While accuracy is a popular metric for model evaluation, it provides only a coarse assessment by considering a single model's score on all dataset items. This paper explores Item Response Theory (IRT), a framework that infers interpretable latent parameters for an ensemble of models and each dataset item, enabling richer evaluation and analysis beyond the single accuracy number. Leveraging IRT, we assess model calibration, select informative data subsets, and demonstrate the usefulness of its latent parameters for analyzing and comparing models and datasets in computer vision.