21 August 2023 in IEEE Transactions on Instrumentation and Measurement
This paper presents a deep data-driven inertial measurement unit (IMU) online calibration (DUET) method that can compensate for the run-time errors of the accelerometer and gyroscope to improve inertial-based odometry. We design a differential error learning strategy based on the kinematic motion model to train the sensor error compensation model. This strategy allows our method to learn IMU sensor errors, such as scale factors, axis-misalignment, and biases, solely from displacement and orientation increments given by external tracking systems. Then during the odometry computation, the trained model leverages the past inertial data to mitigate the sensor errors and thus reduces the integration errors to reflect the odometry state. The experiments conducted on two public visual-inertial datasets show an average of 20% improvement in the position estimation accuracy of visual-inertial odometry, which is comparable to existing learning-based methods with lower operational complexity.
19 April 2011 in Applied Physics B
25 June 2025 in Royal Society Open Science
Researchers are increasingly using machine learning to study physiological markers of emotion. We evaluated the promises and limitations of this approach via a big team science competition. Twelve teams competed to predict self-reported affective experiences using a multi-modal set of peripheral nervous system measures. Models were trained and tested in multiple ways: with data divided by participants, targeted emotion, inductions, and time. In 100% of tests, teams outperformed baseline models that made random predictions. In 46% of tests, teams also outperformed baseline models that relied on the simple average of ratings from training datasets. More notably, results uncovered a methodological challenge: multiplicative constraints on generalizability. Inferences about the accuracy and theoretical implications of machine learning efforts depended not only on their architecture, but also how they were trained, tested, and evaluated. For example, some teams performed better when tested on observations from the same (vs. different) subjects seen during training. Such results could be interpreted as evidence against claims of universality. However, such conclusions would be premature because other teams exhibited the opposite pattern. Taken together, results illustrate how big team science can be leveraged to understand the promises and limitations of machine learning methods in affective science and beyond.
12 January 2026 in IEEE Journal of Biomedical and Health Informatics
Early detection of visual impairments remains a persistent challenge, especially due to the subtle and often unnoticed nature of early-stage symptoms. Recent works have attempted to transition clinical tests to home-based services or develop innovative diagnostic methods, but most approaches remain self-initiated and discrete. In this study, we focused on refractive disorders and explored the feasibility of using electrooculography (EOG) to detect changes in refractive power passively. Thirty-nine participants used optometry trial lenses to simulate different refractive conditions. Participants performed a series of visual tasks while their EOG signals were recorded. We trained classification models to predict simulated refractive power levels relative to baseline visual condition across multiple evaluation settings, including within-subject, temporal generalization, and across-subject scenarios. The findings reveal that refractive power classification models achieve a mean accuracy of $0.950 \pm 0.034$ in within-subject, within-condition scenarios. Within-subject models tested on data from a different time point showed highly variable performance. While some participants achieved promising results, overall accuracy remained low, with a mean of $0.159 \pm 0.285$. We employed three strategies to evaluate the across-subject models. Naive models performed poorly ($0.161 \pm 0.063$) and linear normalization provided limited improvement ($0.175 \pm 0.062$). However, the fine-tuning strategy substantially improved the model's performance ($0.785 \pm 0.123$). EOG signals contain useful information for refractive power classification, particularly in personalized contexts. However, generalizing across time and individuals remains challenging. Overall, this work offers valuable insights for advancing EOG-based systems aimed at passive, real-time monitoring of visual conditions.