This is our official implementation of S⁴M!

[arXiv] [Project]
by Heeji Yoon*, Heeseong Shin*, Eunbeen Hong, Hyunwook Choi, Hansang Cho, Daun Chung, Seungryong Kim†

We propose S⁴M, a semi-supervised instance segmentation framework that effectively leverages capabilities of Segment Anything Model (SAM) through three key components: structural distillation, pseudo-label refinement, and specialized data augmentation.

Please follow installation.

Please follow dataset preperation.

Training is split into two consecutive steps. We provide shell scripts for each stage:

• train_teacher.sh: Pre-train the teacher model
• train_student.sh: Train the semi-supervised student model. Please note that this stage requires a minimum of 2 GPUs.

# First train teacher model
bash ./scripts/cityscapes/train_teacher.sh [NUM_GPUS] [PERCENTAGE] 

# After completing pre-training teacher model, train student model 
bash ./scripts/cityscapes/train_student.sh [NUM_GPUS] [PERCENTAGE] [TEACHER_CKPT]

# evaluation
bash ./scripts/cityscapes/val.sh [NUM_GPUS] [TEACHER_CKPT] SSL.TRAIN_SSL False SSL.EVAL_WHO TEACHER
bash ./scripts/cityscapes/val.sh [NUM_GPUS] [STUDENT_CKPT] SSL.TRAIN_SSL True SSL.EVAL_WHO STUDENT

We want to note that, in case the of Cityscapes, teacher pre-training is configured to run on 2× RTX 3090 GPUs, and semi-supervised student training is configured to run on 4× RTX 3090 GPUs by default.

bash ./scripts/coco/train_teacher.sh [NUM_GPUS] [PERCENTAGE] 

# After completing pre-training teacher model, train student model 
bash ./scripts/coco/train_student.sh [NUM_GPUS] [PERCENTAGE] [TEACHER_CKPT]

# evaluation
bash ./scripts/coco/val.sh [NUM_GPUS] [TEACHER_CKPT] SSL.TRAIN_SSL False SSL.EVAL_WHO TEACHER
bash ./scripts/coco/val.sh [NUM_GPUS] [STUDENT_CKPT] SSL.TRAIN_SSL True SSL.EVAL_WHO STUDENT

For the COCO dataset, teacher pre-training is configured to run on 2× RTX A6000 GPUs, and semi-supervised student training is configured to run on 4× RTX A6000 GPUs by default.

We provide pre-trained teacher and student model checkpoints for Cityscapes and COCO.

Code is largely based on GuidedDistillation and SAM2. We sincerely thank the original authors for their invaluable work and contributions!

If you find this research useful, please consider citing:

@misc{yoon2025s4mboostingsemisupervisedinstance,
      title={S^4M: Boosting Semi-Supervised Instance Segmentation with SAM},
      author={Heeji Yoon and Heeseong Shin and Eunbeen Hong and Hyunwook Choi and Hansang Cho and Daun Jeong and Seungryong Kim},
      year={2025},
      eprint={2504.05301},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.05301},
}