This repository contains the detail procedure to train a model for Pill Detection. Upgradations are still in progress.
Step 1: ANNOTATION
- Collect the dataset and annotate it.
- For annotation I have used "LabelImg" tool.
- Just clone the labelImg repo "https://github.com/heartexlabs/labelImg".
- In "data/predefined_classes.txt" include the desired class name and save with the updates list of classes.
- run labelimg.py
- .xml file will be separately for each image.
Step 2: CONVERSION FROM .xml TO .csv
- run xml_to_csv.py to save the annotated data in .csv format using the command: python xml_to_csv.py
- this will save a csv file with the name "pills_label.csv"
Step 3: SPLITTING THE LABELS
- This will split the pill_labels.csv into train and test
Step 4: TRAINING
- mscoco_label_map.pbtxt is used as label map.
- A training pipeline is provided in training_pipeline.config.
- After training weights are saved in model/frozen_inference_graph.pb
Step 5: TESTING
- For testing, include all the testing images inside test folder.
- On cmd, after directing to the path where "test.py" is situated and run test.py as: python test.py