Основні поняття
A novel subject-based multi-source domain adaptation method that leverages data from multiple subjects (treated as distinct domains) to adapt a deep facial expression recognition model to an unlabeled target subject.
Анотація
The paper proposes a subject-based multi-source domain adaptation (MSDA) method for facial expression recognition (FER) that can effectively leverage data from multiple subjects (treated as distinct domains) to adapt a deep FER model to an unlabeled target subject.
Key highlights:
- Existing FER datasets typically contain data from diverse individuals, but do not emphasize subject-specific variations, which are crucial for developing subject-specific models.
- The proposed subject-based MSDA method treats each subject as a distinct domain and aims to mitigate the domain shift among the source subjects before adapting to the target subject.
- An Augmented Confident Pseudo-Label (ACPL) strategy is introduced to generate reliable pseudo-labels for the target subject, which are then used to train the adapted model.
- Experiments on the BioVid and UNBC-McMaster datasets show that the proposed subject-based MSDA method outperforms source-only and state-of-the-art MSDA approaches, and scales well to a large number of source domains.
- Selecting the most relevant source subjects (domains) is shown to significantly improve the model's performance on the unlabeled target subject.
Статистика
Facial expressions can vary significantly depending on the individual and capture conditions, decreasing the performance of FER systems.
The BioVid dataset consists of 87 subjects with 5 pain levels, and the UNBC-McMaster dataset has 25 subjects with 5 pain intensity levels.
Experiments were conducted with 10, 30, 50, 60, and 77 source subjects adapted to 10 unlabeled target subjects on the BioVid dataset.
Цитати
"Facial expressions vary significantly depending on the individual (e.g., their expressiveness) and capture conditions, which decreases the performance of FER systems."
"Unlike UDA, MSDA methods can leverage multiple source datasets to improve the accuracy and robustness of the target model."