Deep Learning for Detecting Amphoras in Ancient Shipwrecks
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 thecsv
and TensorFlowrecord
files, which were converted fromxml
using the scripts inscripts/preprocessing
. Thelabel_map.pbtxt
file defines the class, which isamphora
in our case.images
contains all the training, validation, and test images in thetrain
,val
, andtest
directories respectively. Thexml
files produced by labelImg are also stored here.models/ssd_resnet50_v1_fpn
stores thetfevents
files for TensorBoard evaluation. Thepipeline.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 thexml
files inimages
tocsv
files inannotations
;preprocessing/generate_tfrecord.py
converts thecsv
files to TensorFlowrecord
files inannotations
.statistics/count_objects.py
counts the number of objects and computes the train: val ratio from thexml
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.