Please excuse our dust!

This code is very much in Beta and needs a through deep clean.

This work has been accepted to TPDL 2022 as Figure and Figure Caption Extraction for Mixed Raster and Vector PDFs: Digitization of Astronomical Literature with OCR Features.

Writing of talk (and, let's be honest, eating of Italian food) took precedence, but expect cleaner code mid-October.

Order

1. change paths in config.py for system
2. ocr_and_image_processing_batch.py -- runs the OCR and pdfmining to get raw data from pages
3. pull_check_makesense.ipynb -- uses a current model to "guess" boxes, and prepares them to check with MakeSense.ai
4. process_annotations_and_generate_features_batch.py -- process annotations after downloading (as a csv file) from MakeSense.ai, generate features as well if you want
5. generate_features_only.py -- generates features in batchs for certain feature sets, saves them in tfrecords format
6. mega_yolo_train_tfrecords.ipynb -- to be run on the cloud (set up for Google Collab), trains the model. Make sure to download weights if not doing all work on Collab.
7. post_processing_tfrecords.py -- post processes results of test dataset (in tfrecords format) from saved weights
8. explore_calculate_metrics.ipynb -- takes in post-process results, calculates various metrics, makes nice plots for talk

For exploring data

• explore_features.ipynb -- look at plots of different features
• explore_post_procesing.ipynb -- explore effects of post-processing steps.

TODO

• for pull_check_makesense.ipynb -- make sure there is an option when there is no model already run
• for process_annotations_and_generate_features_batch.py -- maybe this should just be annotation generation? and then generate_features_only.py is the follow up?
• what file is PDFmining in?