Evaluation metrics and dataset bias in salient object detection

Determine robust evaluation metrics for salient object detection and characterize how dataset bias affects model performance, to enable fair and reliable benchmarking across datasets and methods.

Background

The paper surveys 228 publications on salient object detection, detailing models, datasets, and evaluation measures. It notes that evaluation practices vary (e.g., precision–recall, F-measure, ROC/AUC, MAE) and that choices like binarization thresholds or post-processing can impact reported performance.

The authors also discuss dataset bias (e.g., selection and capture biases) and how biases such as center bias can inflate performance of naive baselines and complicate fair comparisons, motivating improved evaluation and analyses of bias effects.