surya-sci

Changes to regex looking for captions based on examining 50 papers from papers-with-code

Captions are now also detected through heuristics using OCR text (e.g. 'Table X:', 'Figure X:'). Caption bboxes are now merged, like table bboxes.

Minor change (added bbox binary 'touching' detection, with wiggle room)

Merge pull request #148 from VikParuchuri/dev

Update batch sizes and benchmarks