CAST: Contrastive Adaptation and Distillation for Semi-Supervised Instance Segmentation

Paper: arXiv:2505.21904

A three-stage pipeline that leverages limited labels and abundant unlabeled data to train a compact student model via contrastive pixel-wise distillation.

---

---

Overview

CAST is built around three key steps:

1. Adaptation
   Warm up the student on the small labeled set using standard instance-segmentation losses.

2. Contrastive Self-Supervision
   Pull pixel-level features of positive proposals closer in embedding space, while pushing negatives apart.

3. Knowledge Distillation
   Transfer instance-level predictions and refined feature representations from a large teacher to the compact student.

   ---

Debiased Pixel-level Negative Sampling:

---

This design yields a student model that nearly matches its teacher on standard benchmarks despite using far fewer parameters and FLOPs.

Statistics	Performance

---

Citation

If you use this code or ideas from CAST, please cite:

@article{taghavi2025cast,
  title   = {CAST: Contrastive Adaptation and Distillation for Semi-Supervised Instance Segmentation},
  author  = {Taghavi, Pardis and Liu, Tian and Li, Renjie and Langari, Reza and Tu, Zhengzhong},
  journal = {arXiv preprint arXiv:2505.21904},
  year    = {2025}
}