A Novel Multiple-View Adversarial Learning Network for Unsupervised Domain Adaptation Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation action recognition requires labeled target domain samples for many existing methods.
• Key components like feature learning, fusion, and classification are often not integrated into a single end-to-end framework.

Purpose of the Study:

• To propose a novel end-to-end multiple-view adversarial learning network (MAN) for unsupervised domain adaptation action recognition.
• To achieve domain-invariant discriminative feature learning and fuse multi-modal features within a unified framework.

Main Methods:

• Developed a robust spatiotemporal feature extraction network with spatial transform and adaptive intrachannel weight modules for scale invariance.
• Implemented a self-attention mechanism for adaptive fusion of RGB and optical-flow features.
• Utilized a multi-view adversarial learning loss to learn domain-invariant discriminative features.

Main Results:

• The proposed MAN outperforms several state-of-the-art unsupervised domain adaptation action recognition approaches on benchmark datasets.
• Achieved significant improvements, such as 3.7% and 6.1% gains on specific datasets compared to a leading method.
• Constructed and released new benchmark datasets to advance research in this area.

Conclusions:

• The novel MAN framework effectively addresses unsupervised domain adaptation for action recognition.
• The integrated approach of feature extraction, fusion, and adversarial learning yields superior performance.
• The released datasets will foster further research and development in domain adaptation action recognition.