Domain Adaptation for Egocentric Action Recognition
This project focuses on advancing the field of Egocentric Action Recognition through innovative domain adaptation techniques. By leveraging RGB features extracted from the EPIC KITCHEN DATASET with a 3D-inflated network, we explore different domain adaptation methods to enhance generalization across various domain shifts. Our approach combines adversarial learning modules at different temporal aggregation levels with an attentive mechanism and Minimum Class Confusion loss, achieving significant improvements over baseline models.
Main points:
- Developed innovative domain adaptation (DA) methods with PyTorch using RGB features and a 3D-inflated network for temporal video analysis, aiming to improve generalization in egocentric action recognition across varying domain shifts.
- Implemented a novel approach combining adversarial learning with attentive mechanisms and Minimum Class Confusion loss,
improving accuracy of +8% respect to the plain model.
- Demonstrated substantial performance gains in action recognition on the public EPIC-KITCHENS dataset, reaching an average accuracy of 56.8% on unseen data, near to the state of the art.
accuracy of 56.8% on unseen data, near to the state of the art.