When evaluating the model with validation dataset and test dataset, why is Recall@N calculated excluding pano data among the query images?