Training a Student Expert via Semi-Supervised Foundation Model Distillation

Researchers introduced a semi-supervised distillation method for instance segmentation that adapts vision foundation models to a compact student.

The framework uses labeled and unlabeled data, then refines the student to reduce pseudo-label errors and bias.

On Cityscapes and ADE20K, the smaller student model outperformed its larger foundation-model teachers.

The approach suggests high-quality segmentation can be trained more cheaply and deployed more practically with limited labels.