DeepBiRD

A bibliographic reference detection appraoch

Pre-processing

Use the provided python script to preprocess the input images. Once paths to input and output image are set, execute the python file:

python3 pre-process-images.py

Following images show different pre-processing stages:

Dilated Example	Distance Transform Example	Pre-processed Example

Installation

For this project, we used the MaskRCNN implementation provided by Detectron.

• Configure Detectron by following these instructions
• Convert the pre-processed dataset to coco format
• Register dataset with Detectron
• Prepare config file for training
• Train and evaluate model

Fine-tuned Model

The model fine-tuned on BibX and BibLy dataset can be downloaded here.