Balancing Accuracy, Calibration, and Efficiency in Active Learning with Vision Transformers Under Label Noise