tools/apply_visual_semantics.py

#!/usr/bin/env python3
"""Merge manually reviewed four-view semantics into the public metadata indexes."""

from __future__ import annotations

import json
from pathlib import Path


ROOT = Path(__file__).resolve().parents[1]
ANN = ROOT / "annotations"


def read_jsonl(path: Path) -> list[dict]:
    with path.open(encoding="utf-8") as handle:
        return [json.loads(line) for line in handle if line.strip()]


def write_jsonl(path: Path, rows: list[dict]) -> None:
    with path.open("w", encoding="utf-8") as handle:
        for row in rows:
            handle.write(json.dumps(row, ensure_ascii=False, separators=(",", ":")) + "\n")


def summarize_features(annotation: dict) -> str:
    vf = annotation["visual_features"]
    terms: list[str] = []
    if vf["openings"]:
        terms.append("openings: " + "; ".join(
            item if isinstance(item, str) else f"{item.get('count')} × {item.get('type')} ({item.get('location')})"
            for item in vf["openings"]
        ))
    if vf["recesses"]:
        terms.append("recesses: " + "; ".join(vf["recesses"]))
    if vf["protrusions"]:
        terms.append("protrusions: " + "; ".join(vf["protrusions"]))
    if vf["repeated_structures"]:
        terms.append("repetition: " + "; ".join(vf["repeated_structures"]))
    terms.append("symmetry: " + vf["symmetry"])
    return " | ".join(terms)


def build_prompts(sample_id: str, label: str, annotation: dict) -> dict:
    primary = annotation["primary_form"]
    caption_en = annotation["visual_caption_en"]
    caption_zh = annotation["visual_caption_zh"]
    feature_plan = ", ".join(annotation["feature_tree_hint"])
    feature_summary = summarize_features(annotation)
    return {
        "image_to_cad_en": (
            f"Reconstruct {sample_id}, a {primary}, from its four standardized isometric grayscale views "
            f"as an executable CadQuery Python program. Visually verified target: {caption_en} "
            f"Use this feature-order hint where suitable: {feature_plan}. Preserve the verified topology, "
            "counts, symmetry, and relative placement. Include all imports, assign the final valid model "
            "to result, and return Python code only."
        ),
        "image_to_cad_zh": (
            f"根据四张标准等距灰度视图，将 {sample_id} 重建为可执行 CadQuery Python 程序。"
            f"人工视觉核验结果：{caption_zh} 建议特征顺序：{feature_plan}。请保持已核验的拓扑、数量、"
            "对称关系与相对位置；包含全部导入，将最终有效模型赋值给 result，只返回 Python 代码。"
        ),
        "mesh_to_cad_en": (
            f"Convert the normalized clean mesh of {sample_id} into an editable CadQuery model. The object "
            f"is visually identified as a {primary}. Match these reviewed observations: {feature_summary}. "
            f"A useful construction sequence is: {feature_plan}. Include all imports, assign the final valid "
            "model to result, and return Python code only."
        ),
        "caption_en": (
            f"Write a precise geometric caption for {sample_id}. Use the manually reviewed observation as "
            f"ground truth: {caption_en} Report only visible geometry and topology."
        ),
        "caption_zh": (
            f"为 {sample_id} 撰写精确几何描述。以人工逐视图核验结果为真值：{caption_zh} "
            "仅描述可见几何与拓扑。"
        ),
        "pbr_generation_en": (
            f"Create an uncropped photorealistic product render of {sample_id}, preserving its identity as a "
            f"{primary}. The render must retain these reviewed geometric cues: {feature_summary}. Use the "
            "recorded material, ground texture, three-point lighting, and camera parameters."
        ),
        "negative_prompt_en": (
            f"cropped {primary}, incorrect feature count, lost symmetry, omitted reviewed openings or recesses, "
            "extra parts, floating fragments, extreme perspective, text, watermark, logo, low contrast, "
            "overexposure, underexposure"
        ),
    }


def main() -> None:
    annotations = {row["sample_id"]: row for row in read_jsonl(ANN / "visual_semantics.jsonl")}
    metadata = read_jsonl(ANN / "metadata_extended.jsonl")
    if len(metadata) != 100 or set(annotations) != {row["sample_id"] for row in metadata}:
        raise SystemExit("Visual semantics and metadata sample IDs do not match exactly.")

    prompt_rows: list[dict] = []
    for row in metadata:
        sample_id = row["sample_id"]
        visual = annotations[sample_id]
        old_descriptions = row["descriptions"]
        row["descriptions"] = {
            "en": visual["visual_caption_en"],
            "zh": visual["visual_caption_zh"],
            "technical_en": old_descriptions.get("technical_en", old_descriptions["en"]),
            "technical_zh": old_descriptions.get("technical_zh", old_descriptions["zh"]),
            "annotation_method": visual["annotation_method"],
        }
        row["visual_semantics"] = {
            "primary_form": visual["primary_form"],
            "features": visual["visual_features"],
            "feature_tree_hint": visual["feature_tree_hint"],
            "confidence": visual["confidence"],
        }
        row["prompts"] = build_prompts(sample_id, row["label"], visual)
        prompt_rows.append({
            "sample_id": sample_id,
            "label": row["label"],
            "split": row["split"],
            "descriptions": row["descriptions"],
            "visual_semantics": row["visual_semantics"],
            "prompts": row["prompts"],
        })

    write_jsonl(ANN / "metadata_extended.jsonl", metadata)
    write_jsonl(ANN / "prompts.jsonl", prompt_rows)
    print("Applied 100 manually reviewed visual-semantic records.")


if __name__ == "__main__":
    main()