Unsupervised Domain Adaptation for Action Recognition via Self-Ensembling and Conditional Embedding Alignment
Indrajeet Ghosh, Garvit Chugh, Abu Zaher Md Faridee, Nirmalya Roy·October 23, 2024
Summary
µDAR, a novel unsupervised domain adaptation method, addresses cross-user adaptation, limited annotations, and domain discrepancies in wearable human action recognition. It combines consistency regularization, temporal ensembling for robust pseudo-label generation, and conditional distribution alignment using kernel-based class-wise maximum mean discrepancy. This integration results in improved generalization and pseudo-label quality, outperforming six state-of-the-art UDA methods across four benchmark datasets by 4-12% in average macro-F1 score.
Introduction
Background
Explanation of wearable human action recognition challenges
Importance of unsupervised domain adaptation in addressing these challenges
Objective
Aim of µDAR in solving cross-user adaptation, limited annotations, and domain discrepancies
Method
Consistency Regularization
Description of consistency regularization technique
How it enhances model robustness and generalization
Temporal Ensembling
Explanation of temporal ensembling for pseudo-label generation
Benefits in improving pseudo-label quality and model performance
Conditional Distribution Alignment
Use of kernel-based class-wise maximum mean discrepancy for alignment
How it addresses domain discrepancies and enhances adaptation
Integration
How the three components work together to improve µDAR's effectiveness
Evaluation
Benchmark Datasets
Description of the four benchmark datasets used for evaluation
Performance Metrics
Explanation of the metrics used to assess µDAR's performance
Results
Comparison of µDAR against six state-of-the-art UDA methods
Average macro-F1 score improvement across the four datasets
Conclusion
Summary of µDAR's contributions
Future directions and potential applications
Basic info
papers
computer vision and pattern recognition
artificial intelligence
Advanced features