/amphora_object_detection

Deep Learning for Detecting Amphoras in Ancient Shipwrecks

Primary LanguageJupyter NotebookMIT LicenseMIT

Deep Learning for Detecting Amphoras in Ancient Shipwrecks

Open In Colab

This repository contains the code, the trained model, and the thesis for "Deep Learning for Detecting Amphoras in Ancient Shipwrecks". The simplest way to view the code and the result is through the Open in Colab button at the top.

Project Structure

  • annotations stores the csv and TensorFlow record files, which were converted from xml using the scripts in scripts/preprocessing. The label_map.pbtxt file defines the class, which is amphora in our case.
  • images contains all the training, validation, and test images in the train, val, and test directories respectively. The xml files produced by labelImg are also stored here.
  • models/ssd_resnet50_v1_fpn stores the tfevents files for TensorBoard evaluation. The pipeline.config file is the configuration file used for training.
  • pretrained_models/ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync_2018_07_03 is the pre-trained model obtained from the TensorFlow 1 Detection Model Zoo.
  • scripts stores the Python scripts. preprocessing/xml_to_csv.py converts the xml files in images to csv files in annotations; preprocessing/generate_tfrecord.py converts the csv files to TensorFlow record files in annotations. statistics/count_objects.py counts the number of objects and computes the train: val ratio from the xml files.
  • thesis has the complete thesis write-up in LaTeX. Tianyao_Chen_bachelor_thesis.pdf is the main PDF file.
  • trained_models/ssd_resnet50_v1_fpn.pb has the exported trained model for inference from this study.

References