satellite_detection

• the environment setup guide can be found here

• The entire dataset was taken from SPEED's train split (speed/images/train)

• Ground truth data (bbox.json) was taken from here, then converted to a .csv

• I split the dataset into train/test using split_train_test.py

• tfrecord files for the train and test set can be generated using generate_tfrecord.py

• Optionally, move images in the test split into its own folder using extract_test_set.py

• The training configs and checkpoints for each model are in their respective folders

  • /mobiledet (PTQ)
  • /mobiledet_qat
  • /spaghettinet_qat

• The tflite graphs and files for each model are in their respective folders in /tflite

  • the full integer quantised & Edge TPU compiled files are in /tflite/full_int_quant

• Training was executed through the retrain_od_gpu script.