ModelZoo: web platform to collect data and share models, together with a new paradigm for pre-training and fine-tuning animal pose models.
Improve the generalisation ability of models for animal pose estimation.
Neuro Sense: hardware-constrained human activity recognition, activity embeddings construction and heart rate estimation
Learning activity representation and estimating heart rate using PPG and IMU data from the smartwatch. Activity encoding was trained with self-supervised contrastive learning. Our heart rate estimation model archives mean absolute error 5 bpm and mean absolute percentage error only 4%. Because of the small number of params (5M MAC operations only) tiny AI chip with implemented model consumes only 0.1 mW.
Inertial navigation and activity recognition on smartphones
Developed model for inertial navigation and activity classification on smartphones based on IMU data. Created new algorithm for gyroscope bias correction based on smart quaternions correction. Archived 8.9 RMSE and 6.0 MAPE in trajectory estimation task.
Dehydratation estimation via smartwatch
Built model for sweat loss estimation during workout via Samsung Galaxy Watch 4 (implemented feature) for Samsung Research based on BMI, age, weight and accelerometer and gyroscope data.
Modern Methods for Quantifying Behavior
Last updated on Mar 26, 2023
0 min read