My project is title Bridge, it is an AI powered sign language interpreter, and this week I have created my first model after starting this course

For this week, I have mainly worked on creating the categorization model for my software. This process took a lot of time and effort, however, i have not found any success in training a meaningful model.

The logistics of training the model is as follows:

• Data Collection
  • Collection of data that is consistent and reliable
• Data Preparation
  • Making the data in a format that the machine learning model can understand
  • Separating Learning data and testing data
• Architecture Design
  • Selecting layers to use in the Neural Networks
• Model Fitting
  • Actual Modelling of the data
• Model Assessment
  • Testing numerically how accurate the created model was

Data Collection

This was what I left of last week, I created a system where the datapoints in the body will be recorded if I press the keyboard so that I can record whenever convenient.

I have finished this program however due to how OpenCV was created, every time the keyboard is pressed, it pauses the video stream. After debugging, I have realized that data is still being stored even though the video is not moving meaning that the webcam is still recording. I wanted to continue this but realized that I cannot record without using my other hand to press the spacebar. and most of the gestures needed hands.

The effort put to code this didn't go to waste as I used the same threading logic in my other implementation.

My implementation was now changed where if the hands are visible on the camera, then it will start recording the data. This is useful as the recording immediately stops when I stop using my hands. This worked so I proceeded to record all of the words. Despite this, it eventually became a huge problem.

The way I recorded words was that for each gesture, I repeated it 50 times. I did this across 21 individual words. I realized that the model would pause and not record my hands in weird situations or lightings, so what happens is that the recordings get separated in between and thus unusable.

What I did to fix this was to create a separate script that deletes the whole folder if the inputs are less than a specific number.