prosperious's picture
Add 1000 files (OCR_INDRA/.gitignore ...)
f9bc0c1 verified
Raw History Blame Contribute Delete
3.17 kB
"""Render a synthetic Devanagari (Nepali) test page with known ground truth.
Each GT line is drawn with PIL/raqm (proper shaping via HarfBuzz) at a known
position, so we get both the exact text and per-line boxes. This exists only
for smoke tests of the Indra pipeline on this machine: real evaluation happens
on the remote machine's frozen sets. The font (OFL) and the page here are test
assets, not eval data.
Usage:
PYTHONPATH=/tmp/opencode/indra-lib python3 make_synth_page.py [outdir]
"""
import json
import sys
from pathlib import Path
from PIL import Image, ImageDraw, ImageFont
FONT = Path(__file__).parent / "testdata" / "NotoSansDevanagari.ttf"
# Mixed Nepali coverage: initial/final matras, conjuncts, nukta, danda,
# digits, ॐ, ष/ृ, and a Latin word (vocab contains Latin too).
LINES = [
"नेपाल भाषा ऱूपमा",
"क्षरत्र ज्ञान तन्त्र पुनर्निर्माण",
"ॐ नमः शिवाय ।",
"उपयोगकर्ता सङ्केत द्वोपक स्मृति",
"क१२३४५६७८९० ख२०२६ ग₹।",
"फलमूल असन्न हृदय ऋतु",
"नेपाल काठमाडौँ बराहनगर ॥",
"कितावबाट गर्दैछु भव्य यात्रा",
"५६ सालदेखि ४८ जना ।",
"स्वतन्त्र बृहत् जीवनसाथी",
]
def main(argv=None):
out = Path(argv[0] if argv else ".").resolve()
out.mkdir(parents=True, exist_ok=True)
Wpx, Hpx = 1240, 1754 # ~A4 @150dpi
size = 32 # px; run small: big pages skew the smoke test toward slow det
font = ImageFont.truetype(str(FONT), size)
img = Image.new("RGB", (Wpx, Hpx), "white")
d = ImageDraw.Draw(img)
gt = []
y = 90
for i, text in enumerate(LINES):
d.text((60, y), text, font=font, fill="black")
bb = d.textbbox((60, y), text, font=font)
gt.append(
{
"i": i,
"text": text,
"box": [float(bb[0]), float(bb[1]), float(bb[2]), float(bb[3])],
}
)
y += int(size * 2.0)
page = out / "synth_nepali_page.png"
img.save(page)
gt_file = out / "synth_nepali_gt.jsonl"
gt_file.write_text(json.dumps({"file": page.name, "lines": gt}) + "\n")
# predict-schema record (image/lines/text) for `cer`
rec_file = out / "synth_nepali_record.jsonl"
rec_file.write_text(
json.dumps(
{
"image": str(page),
"page": 1,
"w": Wpx,
"h": Hpx,
"lines": [{"box": g["box"], "text": g["text"], "conf": 1.0} for g in gt],
"text": "\n".join(g["text"] for g in gt),
},
ensure_ascii=False,
)
+ "\n",
encoding="utf-8",
)
print(f"image: {page}\ngt: {gt_file}\nrecord: {rec_file}")
for g in gt:
print(f" {g['i']:2d} {g['text']}")
if __name__ == "__main__":
main(sys.argv[1:])