Fig. 2

An overview of mosaic-based soft labeling with curriculum learning. Our curriculum learning is consisted of two stages. a In stage 1, a conventional training is performed using an original dataset with the weight balancing method. Transfer learning, data augmentation, and fine-tuning was also used during training. b Using the trained model in Stage 1, a mosaic dataset and the corresponding soft labels can be obtained as described in Fig. 1. c In stage 2, the final model is trained with the original dataset and a newly added mosaic-based soft labeling dataset