Ignore zero-height and zero-width bounding boxes.

Signed-off-by: Benjamin Hahn <[email protected]>

Author: b-hahn

docling_core/types/doc/document.py

@@ -5347,12 +5347,17 @@ def load_from_doctags(  # noqa: C901
         def extract_bounding_box(text_chunk: str) -> Optional[BoundingBox]:
             """Extract <loc_...> coords from the chunk, normalized by / 500."""
             coords = re.findall(r"<loc_(\d+)>", text_chunk)
-            if len(coords) > 4:
+            if len(coords) < 4:
+                return None
+            else:
                 coords = coords[:4]
-            if len(coords) == 4:
                 l, t, r, b = map(float, coords)
+                eps = 1e-3
+                # Ignore bounding boxes with width or height of <1e-3, including cases where l>r or t>b.
+                if r - l < eps or b - t < eps:
+                    return None
+
                 return BoundingBox(l=l / 500, t=t / 500, r=r / 500, b=b / 500)
-            return None
 
         def extract_inner_text(text_chunk: str) -> str:
             """Strip all <...> tags inside the chunk to get the raw text content."""

test/data/doc/defect_bbox_page.dt

@@ -0,0 +1,4 @@
+Assistant: <doctag>
+<text><loc_68><loc_447><loc_445><loc_447>This is valid text with a zero-height bounding box.</text>
+<text><loc_68><loc_447><loc_32><loc_499>This is valid text with a negative-width bounding box.</text>
+</doctag><end_of_utterance>

test/data/doc/defect_bbox_page.dt.json

@@ -0,0 +1,66 @@
+{
+  "body": {
+    "children": [
+      {
+        "$ref": "#/texts/0"
+      },
+      {
+        "$ref": "#/texts/1"
+      }
+    ],
+    "content_layer": "body",
+    "label": "unspecified",
+    "name": "_root_",
+    "self_ref": "#/body"
+  },
+  "form_items": [],
+  "furniture": {
+    "children": [],
+    "content_layer": "furniture",
+    "label": "unspecified",
+    "name": "_root_",
+    "self_ref": "#/furniture"
+  },
+  "groups": [],
+  "key_value_items": [],
+  "name": "Document",
+  "pages": {
+    "1": {
+      "page_no": 1,
+      "size": {
+        "height": 1.0,
+        "width": 1.0
+      }
+    }
+  },
+  "pictures": [],
+  "schema_name": "DoclingDocument",
+  "tables": [],
+  "texts": [
+    {
+      "children": [],
+      "content_layer": "body",
+      "label": "text",
+      "orig": "This is valid text with a zero-height bounding box.",
+      "parent": {
+        "$ref": "#/body"
+      },
+      "prov": [],
+      "self_ref": "#/texts/0",
+      "text": "This is valid text with a zero-height bounding box."
+    },
+    {
+      "children": [],
+      "content_layer": "body",
+      "label": "text",
+      "orig": "This is valid text with a negative-width bounding box.",
+      "parent": {
+        "$ref": "#/body"
+      },
+      "prov": [],
+      "self_ref": "#/texts/1",
+      "text": "This is valid text with a negative-width bounding box."
+    }
+  ],
+  "version": "1.8.0"
+}

test/test_doctags_load.py

@@ -168,3 +168,17 @@ def test_doctags_inline():
         exp_file=exp,
         actual=deser_doc.export_to_dict(),
     )
+
+
+def test_doctags_handle_defect_bbox():
+
+    doctags_doc = DocTagsDocument.from_doctags_and_image_pairs(
+        [Path("test/data/doc/defect_bbox_page.dt")], None
+    )
+
+    doc = DoclingDocument.load_from_doctags(doctags_doc)
+    exp = "test/data/doc/defect_bbox_page.dt.json"
+    verify(
+        exp_file=exp,
+        actual=doc.export_to_dict(),
+    )