# /// script # requires-python = ">=3.10" # dependencies = [] # /// """Export VietFinTab to a YOLO detection dataset: classes row, column, cell, table. The source has no plain cell boxes, so cells are derived: every row x column intersection not covered by a spanning cell, plus the spanning cells themselves. Usage: uv run scripts/export_yolo.py [--root .] [--out VietFinTab-reconstruction] [--val 0.1] [--seed 0] """ import argparse import json import random import shutil import xml.etree.ElementTree as ET from pathlib import Path NAMES = ["row", "column", "cell", "table"] def boxes(xml_path): r = ET.parse(xml_path).getroot() W, H = int(r.findtext("size/width")), int(r.findtext("size/height")) by = {} for o in r.iter("object"): b = o.find("bndbox") by.setdefault(o.findtext("name"), []).append( tuple(float(b.findtext(k)) for k in ("xmin", "ymin", "xmax", "ymax"))) return (W, H, by.get("table row", []), by.get("table column", []), by.get("table spanning cell", []), by.get("table", [])) def cells(rows, cols, spans): out = list(spans) for _, y0, _, y1 in rows: for x0, _, x1, _ in cols: cx, cy = (x0 + x1) / 2, (y0 + y1) / 2 if not any(a <= cx <= c and b <= cy <= d for a, b, c, d in spans): out.append((x0, y0, x1, y1)) return out def yolo_lines(W, H, groups): lines = [] for cls, bs in enumerate(groups): for x0, y0, x1, y1 in bs: x0, x1, y0, y1 = max(x0, 0), min(x1, W), max(y0, 0), min(y1, H) if x1 <= x0 or y1 <= y0: continue # degenerate after clipping lines.append(f"{cls} {(x0 + x1) / 2 / W:.6f} {(y0 + y1) / 2 / H:.6f} " f"{(x1 - x0) / W:.6f} {(y1 - y0) / H:.6f}") return lines def export(root, out, val, seed): meta = [json.loads(l) for l in (root / "metadata.jsonl").read_text().splitlines()] random.Random(seed).shuffle(meta) n_val = round(len(meta) * val) for i, m in enumerate(meta): split = "val" if i < n_val else "train" W, H, rows, cols, spans, table = boxes(root / m["structure_file"]) for kind in ("images", "labels"): (out / kind / split).mkdir(parents=True, exist_ok=True) shutil.copy2(root / m["file_name"], out / "images" / split / f"{m['id']}.jpg") (out / "labels" / split / f"{m['id']}.txt").write_text( "\n".join(yolo_lines(W, H, [rows, cols, cells(rows, cols, spans), table])) + "\n") names = "\n".join(f" {i}: {n}" for i, n in enumerate(NAMES)) (out / "data.yaml").write_text( f"path: {out.resolve()}\ntrain: images/train\nval: images/val\nnames:\n{names}\n") print(f"{len(meta) - n_val} train / {n_val} val -> {out}/") if __name__ == "__main__": ap = argparse.ArgumentParser() ap.add_argument("--root", type=Path, default=Path(__file__).resolve().parent.parent) ap.add_argument("--out", type=Path, default=Path("datasets/VietFinTab-reconstruction-yolo")) ap.add_argument("--val", type=float, default=0.1) ap.add_argument("--seed", type=int, default=0) a = ap.parse_args() export(a.root, a.out, a.val, a.seed)