Notes. The localization is performed on predictions that have been renormalized to be within the same range as the original sky model. The model configurations vary in terms of normalization power used during training, and also in terms of the aggregation method employed. For the normalization with high root powers γ, the optimal aggregation method is detect-aggregate. This approach offers a good trade-off between purity and completeness, resulting in the best F1-score. For the lower γ, the best approach turned out to be aggregate-detect. We compare the results with Mohan & Rafferty (2015; PyBDSF) and Taran et al. (2023). We run the Photutils algorithm directly on sky models to quantify the error coming from this step, denoted as Photutils localization. Italic font indicates the optimal metric value across all conducted experiments.