To search, Click below search items. |
| All Published Papers Search Service |
|
|
|
Title |
Transfer Learning for Cross-Domain Eye Gaze Classification: Quantifying Domain Gap and Fine-Tuning Effectiveness |
|
Author |
Rumana Ferdushi |
| Citation |
Vol. 26 No. 2 pp. 39-46 |
|
Abstract |
Eye tracking systems are essential for human-computer interaction, assistive technology, and driver monitoring; however, the domain shift makes it difficult to deploy models across various cameras and locations. This study investigates transfer learning as a viable solution to cross-domain eye gaze classification. We start by calculating the domain gap: a CNN trained on a source dataset achieves 99.76% in-domain accuracy, but when tested on a target dataset with a synthetic domain shift (14.84% gap), the accuracy reduces to 84.9%. Next, we use transfer learning through fine-tuning on the target domain, regaining 58.8% of lost performance and attaining 93.7% accuracy. Our results demonstrate that transfer learning successfully overcomes domain shifts in eye classification, allowing for reliable deployment in a variety of visual scenarios. Importantly, practitioners can achieve near-in-domain performance with minimal target-domain annotation, significantly reducing deployment costs. |
|
Keywords |
Transfer learning, domain adaptation, eye tracking, gaze classification, domain generalization, convolutional neural networks, biomedical signal processing. |
|
URL |
