MambaGaze: AI Framework for Cognitive Load Assessment from Eye-Tracking
A group of researchers has unveiled MambaGaze, a groundbreaking framework that assesses cognitive load in real-time using eye-tracking data. It tackles challenges like data loss from blinks and inaccuracies in tracking. The framework incorporates XMD encoding to manage data uncertainty and employs the bidirectional Mamba-2 model to understand long-term temporal relationships. In trials with the CLARE and CL-Drive datasets, MambaGaze achieved accuracy rates of 76.8% and 73.1%, outperforming CNN, Transformer, ResNet, and VGG by 4 to 12 percentage points. Moreover, tests on NVIDIA Jetson show its promise for critical safety applications, such as monitoring driver focus and supporting automated flight decks.
Key facts
- MambaGaze is a framework for cognitive load assessment from eye-tracking data.
- It uses XMD encoding to handle missing data from blinks and tracking failures.
- Bidirectional Mamba-2 captures temporal dependencies with linear complexity.
- Achieved 76.8% accuracy on CLARE dataset and 73.1% on CL-Drive dataset.
- Outperforms CNN, Transformer, ResNet, and VGG baselines by 4-12 percentage points.
- Edge deployment benchmarks conducted on NVIDIA Jetson.
- Potential applications include driver vigilance monitoring and automated flight deck assistance.
- The paper is available on arXiv under ID 2605.22775.
Entities
Institutions
- arXiv
- NVIDIA