Tensorflow Lite Pipeline

check model_maker_object_detection.ipynb for a in depth look at how the model is trained.

Requirements

• CUDA Version 11.x download
• cuDNN Version 8.1 download
• ffmpeg download
• Tensorflow 2.5.0
• Python 3.6 -> 3.9

Create Virtual Enviroment

python -m venv venv

# cd into the scripts directory
cd venv/Scripts

# activate the virtual enviroment
activate

# Change to the parent directory where all the scripts are
cd ..

Install Dependencies

pip install tensorflow==2.5.0
pip install --use-deprecated=legacy-resolver tflite_model_maker
pip install pycocotools
pip install numpy
pip install opencv-python

Converting video into frames (Optional)

The video converter script has two flags --video to specify the video location and --frame will tell ffmpeg how many frames it should generate per second three is a good median.

py convert.video.py --video ./dataset/videos/one.mp4 --dir-name images-2

Mix data from the different folders of data

mixing the data will combine all the annotations and images listed in the config.json file into annotations and images folders for create-data.py to process.

py dataset.mixer.py

Create dataset

create-dataset.py will take all the files in the images and annotations folders and randomly copy them to the test and train folder's for tensorflow to use.

When making your own dataset use a minimum of 233 images for training to label data using labelimg is a good choice.

py create-dataset.py

Train the model

Model training collects all the annotations and images and passes them to tensorflow to start training the object detection model.

py new.py