fix(layout,table): use default v4 backend for e2e OCR test and fix table structure detection

Signed-off-by: Clément Doumouro <[email protected]>

Author: ClemDoum

docling/models/layout_model.py

@@ -4,7 +4,7 @@
 from collections.abc import Iterable
 from copy import deepcopy
 from pathlib import Path
-from typing import List, Optional, Union
+from typing import List, Optional, Union, cast
 
 import numpy as np
 from docling_core.types.doc import DocItemLabel
@@ -195,7 +195,7 @@ def __call__(
                 continue
 
             page_predictions = batch_predictions[valid_page_idx]
-            page_image = valid_page_images[valid_page_idx]
+            page_image = valid_page_images[valid_page_idx]  # type: ignore[assignment]
             page_orientation = valid_page_orientations[valid_page_idx]
             valid_page_idx += 1
 
@@ -207,23 +207,25 @@ def __call__(
                 bbox = BoundingBox.model_validate(pred_item)
                 if page_orientation:
                     bbox = rotate_bounding_box(
-                        bbox, page_orientation, page_image.size
+                        bbox,
+                        page_orientation,
+                        page_image.size,  # type: ignore[union-attr]
                     ).to_bounding_box()
                 cluster = Cluster(
                     id=ix,
-                     label=label,
-                     confidence=pred_item["confidence"],
-                     bbox=bbox,
+                    label=label,
+                    confidence=pred_item["confidence"],
+                    bbox=bbox,
                     cells=[],
                 )
                 clusters.append(cluster)
 
             if settings.debug.visualize_raw_layout:
                 self.draw_clusters_and_cells_side_by_side(
                     conv_res,
-                            page,
-                            clusters,
-                            mode_prefix="raw",
+                    page,
+                    clusters,
+                    mode_prefix="raw",
                 )
 
             # Apply postprocessing
@@ -253,9 +255,9 @@ def __call__(
             if settings.debug.visualize_layout:
                 self.draw_clusters_and_cells_side_by_side(
                     conv_res,
-                        page,
-                        processed_clusters,
-                        mode_prefix="postprocessed",
+                    page,
+                    processed_clusters,
+                    mode_prefix="postprocessed",
                 )
 
             yield page

docling/models/table_structure_model.py

@@ -202,6 +202,10 @@ def __call__(
                         continue
                     # Rotate and scale table image
                     page_im = cast(Image, page.get_image())
+                    original_scaled_page_size = (
+                        int(page_im.size[0] * self.scale),
+                        int(page_im.size[1] * self.scale),
+                    )
                     scaled_page_im: Image = cast(
                         Image, page.get_image(scale=self.scale)
                     )
@@ -258,10 +262,12 @@ def __call__(
                                             scale=self.scale
                                         )
                                     )
+                                    # _rotate_bbox expects the size of the image in
+                                    # which the bbox was found
                                     new_bbox = _rotate_bbox(
                                         new_cell.to_bounding_box(),
-                                        orientation=cells_orientation,
-                                        im_size=scaled_page_im.size,
+                                        orientation=-cells_orientation,
+                                        im_size=original_scaled_page_size,
                                     ).model_dump()
                                     tokens.append(
                                         {
@@ -286,7 +292,11 @@ def __call__(
                                         the_bbox
                                     )
                                     element["bbox"]["token"] = text_piece
-
+                                element["bbox"] = _rotate_bbox(
+                                    BoundingBox.model_validate(element["bbox"]),
+                                    orientation=cells_orientation,
+                                    im_size=scaled_page_im.size,
+                                ).model_dump()
                                 tc = TableCell.model_validate(element)
                                 if tc.bbox is not None:
                                     tc.bbox = tc.bbox.scaled(1 / self.scale)