Learning to Recognize Actions When Labels and Light Are Scarce: Semi-Supervised, Cross-Domain, and Dark-Video Methods

This thesis tackles two obstacles in deploying human action recognition in real-world video: reliance on expensive densely labeled clips and poor performance in low-light conditions. First, SpIVA transfers supervision from cheaper labeled images to video models, enabling action learning from heterogeneous visual sources. Second, ActNetFormer is a semi-supervised framework where complementary video models teach each other using abundant unlabeled footage. Third, ActLumos targets dark/low-light videos by combining original and Retinex-enhanced views, then distilling them into an efficient single-stream model. Experiments on standard and low-light benchmarks show consistent gains, improving robustness for safety-critical applications.