Fixed a rare bug in OCR system where self-intersecting bounding boxes…

… caused crashes during multi-pass OCR. Implemented a quick fix in bbox_intersection_area_ratio function to return 0.0 when polygons are invalid due to self-intersection.

CHANGELOG.md

@@ -10,7 +10,7 @@ The version numbers are according to [Semantic Versioning](http://semver.org/).
 
 ### Fixed
 - Adds forced conversion to RGB in pillow before sending data to OpenCV to fix a possible bug in Studio
-
+- Fixes a rare bug where self-intersecting bounding boxes caused the OCR system to crash when using multi-pass OCR
 ### Removed
 
 

ocr_wrapper/bbox_utils.py

@@ -15,6 +15,9 @@ def bbox_intersection_area_ratio(bb1: BBox, bb2: BBox) -> float:
     """
     self_poly = bb1.get_shapely_polygon()
     that_poly = bb2.get_shapely_polygon()
+    # Sometimes the polygons are invalid (usually because of self-intersection), in which case we return 0.0. We should have a closer look why this happens, but for now it seems to be a rare occurence and this is a quick fix that doesn't seem to affect the results much.
+    if not self_poly.is_valid or not that_poly.is_valid:
+        return 0.0
     if self_poly.intersects(that_poly):
         inter_poly = self_poly.intersection(that_poly)
         return inter_poly.area / self_poly.area