Spatio-Temporal Learning from Limited Labels for Video Action Detection

Explore spatio-temporal learning from limited labels in this invited talk from the ECCV ATLAS Tutorial. Delve into video action detection, semi-supervised learning techniques, and temporal coherence. Examine the adaptive proximity-aware uncertainty approach and its modules. Analyze training objectives, compare supervised and semi-supervised methods, and investigate ablation studies. Discover frame selection methods, loss formulations, and key findings in experiments. Gain insights into addressing high annotation costs and improving generalization in video understanding tasks.

Intro
Outline
Introduction
Motivation
Video action detection
Semi-supervised learning
Temporal coherence
Temporal gradient
Modules walk-through
Training objectives
Comparison with supervised
Comparison with semi-supervised
Ablations - break it down
Additional data
Generalization
Conclusion
High annotation cost
Adaptive Proximity-aware Uncertainty A
Overall approach
Experiments
Comparisons
Frame selection methods
Qualitative frame selection analysis
Analyzing loss formulations
Key findings