(1) init_pretrained_model: Model pretrained with default settings of DeeplabV2 on the seeds mask of M&G+ from SDS.

(2) bcm_model: Model finetuned with BCM supervision.

(3) fr_model: Final model finetuned under FR loss.

(4) pseudo_semantic_labels_trainaug: the final generated pseudo labels of train_aug set, which can be used for further training with any fully-supervised models.

(5) pseudo_inst_labels_trainaug_val: the COCO-like instance pseudo labels of train_aug and val set, which can be used for further training with instance segmentation models like Mask RCNN in Detectron2 and MMDetction.