The Staged Knowledge Distillation in Video Classification: Harmonizing Student Progress by a Complementary Weakly Supervised Framework

Chao Wang and Zheng Tang

IEEE T-CSVT

Summary

This framework studies knowledge distillation for video classification under limited supervision. It combines substage-based teacher-student alignment, progressive cascade training, pseudo-label refinement, and a feature-distribution loss. These components address differences between teacher and student capacity while transferring complementary information across network stages. Experiments on real and simulated video datasets evaluate the resulting classification performance.

Citation

@article{Wang24SKD,
  author = {Chao Wang and Zheng Tang},
  title = {The staged knowledge distillation in video classification: {H}armonizing student progress by a complementary weakly supervised framework},
  journal = {T-CSVT},
  volume = {34},
  number = {8},
  pages = {6646--6660},
  year = {2024}
}