Toward Auto-evaluation with Confidence-based Category Relation-aware Regression

Auto-evaluation aims to automatically evaluate a trained model on any test dataset without human annotations. Most existing methods utilize global statistics of features extracted by the model as the representation of a dataset. This not only ignores the influence of the classification head but also loses category-wise confusion information of the model. However, ratios of instances assigned to different categories together with their confidences reflect how many instances in which categories are difficult for the model to classify, which contain significant indicators for both overall and category-wise performances. In this paper, we propose a Confidence-based Category Relation-aware Regression ($C^2R^2$) method. $C^2R^2$ divides all instances in a meta-set into different categories according to their classification confidences by the model to be evaluated and extracts the global representation from the statistics of confidences of all categories. For each specific category, $C^2R^2$ encodes its local confusion relations to other categories into a local representation. The overall and category-wise performances are regressed from global and local representations, respectively. Extensive experiments show the effectiveness of our method.