The code for this project was created by Pierluigi Ferrari in his Github repository keras-retinanet. The project was copied and adapted for this assignment.

RetinaNet was introduced in the paper Focal Loss for Dense Object Detection.

• Python 3.6
• TensorFlow 2.4.0
• Keras 2.4.3
• OpenCV

Exported environment: environment.yml

• Install all dependencies from the environment.yml file (I did not use Jupyter Notebook due to issues with incompatibility of Tensorflow and Jupyter Notebook).
• Download pretrained model Google Drive and put it into main file.
• Put files from keras-retinanet to kerasretinanet file.
• (optional) In a case you want to train model on your own dataset:
  • train model: python kerasretinanet/keras_retinanet/bin/train.py --freeze-backbone --random-transform --weights snapshots/_pretrained_model.h5 --batch-size 8 --steps 500 --epochs 5 csv annotations.csv classes.csv
  • convert model: python kerasretinanet/keras_retinanet/bin/convert_model.py snapshots/resnet50_csv_10.h5 snapshots/inference/resnet50_csv_10.h5
• Run get_results.py to get prediction on the image. You need to add your record to df_test_metrics.csv.
• Run get_metrics.py to get metrics for the model. I used IoU metric (see iou.py script for more information).
• evaluate model: python kerasretinanet/keras_retinanet/bin/evaluate.py csv df_test.csv classes.csv snapshots/inference/resnet50_csv_10.h5