Author: yunjiaz

This week, I worked on model building, training optimization, and dataset processing this week. In order to begin multi-class classification, I first cleaned and prepared the motion data, applied sequence padding, and converted gesture labels into one-hot encoding.

I experimented with three distinct architectures for model development: CNN, RNN, and LSFM. While the RNN (Bidirectional LSTM) assisted in tracking motion across time, the CNN was utilized to record spatial patterns. Both strategies were used in the LSFM model to increase precision and effectiveness.

I make progress to adjust hyperparameters such as learning rates, dropout rates, epoch counts, and batch size. I implemented model checkpointing and early halting to avoid overfitting.

I am on track with the project timeline. Next week, I will fine-tune the models by adjusting kernel sizes, units, and regularization. I will also run more evaluations and start real-time testing to ensure the model works smoothly with the wand system.

This week, I focused on data collection for the gesture recognition model in our wand project. Specifically, I collected data for six different gestures. Specifically,  I recorded 50 samples for each gesture. I manually recorded and labeled each set to ensure the data is consistent. In addition to data collection, I began preliminary data preprocessing, including normalizing the motion data inputs and organizing the dataset into a format compatible with our CNN training pipeline. I also spent time reviewing related research on gesture recognition to identify potential improvements for our model architecture.

My progress is on schedule.

Next week, I plan to work on

1. Develop and implement the initial version of the CNN model for gesture recognition.
2. Conduct initial training and testing of the model to evaluate baseline performance.