[datasets] fix zero-dimension synthetic image for blank glyphs (#2016)

doctr/datasets/generator/base.py

@@ -15,6 +15,11 @@
 
 from ..datasets import AbstractDataset
 
+# Minimum size (in pixels) of a synthesized image dimension. Whitespace and glyphs a font
+# renders blank yield a zero-height/width bounding box; clamping avoids a 0-dimension image
+# that crashes the downstream resize during training (#2016).
+MIN_SYNTH_IMG_DIM = 1
+
 
 def synthesize_text_img(
     text: str,
@@ -41,7 +46,8 @@ def synthesize_text_img(
     font = get_font(font_family, font_size)
     left, top, right, bottom = font.getbbox(text)
     text_w, text_h = right - left, bottom - top
-    h, w = int(round(1.3 * text_h)), int(round(1.1 * text_w))
+    h = max(int(round(1.3 * text_h)), MIN_SYNTH_IMG_DIM)
+    w = max(int(round(1.1 * text_w)), MIN_SYNTH_IMG_DIM)
     # If single letter, make the image square, otherwise expand to meet the text size
     img_size = (h, w) if len(text) > 1 else (max(h, w), max(h, w))
 

tests/common/test_datasets.py

@@ -4,6 +4,17 @@
 import pytest
 
 from doctr import datasets
+from doctr.datasets.generator.base import synthesize_text_img
+
+
+@pytest.mark.parametrize("text", ["  ", "   "])
+def test_synthesize_text_img_never_degenerate(text):
+    # A font renders whitespace (and glyphs it lacks) with a zero-height bounding box.
+    # This used to produce a 0-dimension image that crashes the recognition resize
+    # during training (#2016). The synthesized image must always be at least 1x1.
+    img = synthesize_text_img(text)
+    assert img.height >= 1
+    assert img.width >= 1
 
 
 def test_visiondataset():