ID 2516: Improve IMU-Based Handwriting Recognition with Data Generation

Background

Our model recognizes handwritten words and sentences character by character, classifying each time-series sequence frame into a specific character/symbol category (https://arxiv.org/abs/2502.20954) while the performance of the approach is still limited by the dataset. Data Generation is now playing an important role in Computer Vision field and many works have shown that using generated image for training can significantly improve the performance in downstream task. In this project, we will apply various data generation methods to generate IMU sequences for given texts for training to investigate and investigate whether the generated data can improve the performance on handwriting recognition.

Tasks

• Implement various data generation methods for time-series data.
• Generate IMU sequences from given texts.
• Train IMU-based handwriting recognition models with IMU
• Compare the performance with the baseline
• Support data collection

Requirements

• Proficiency in Python and PyTorch
• Knowledge in related field, e.g. time-series analysis, data generation.
• Speak English and German to support data collection (optional)

Supervisors

Jindong Li

Researcher & PhD Candidate