detectpose.py

from __future__ import annotations

import argparse
import json
import sys
from pathlib import Path
from typing import Optional

import cv2
import numpy as np

ROOT = Path(__file__).resolve().parent
APM = ROOT / "automation_pose_mask"
if str(APM) not in sys.path:
    sys.path.insert(0, str(APM))

from automation_pose_mask.openpose import OpenposeDetector  # noqa: E402


def openpose_to_equifashion_candidate(candidate: list, subset: list) -> list[list[float]]:
    """

    Convert OpenPose (candidate/subset) to EquiFashion-style candidate list.

    Output: list of [x, y, conf, joint_index] for joint_index 0..17 (COCO-18 body).

    """
    cand = np.asarray(candidate, dtype=np.float64)
    if cand.size == 0 or cand.ndim != 2:
        return [[0.0, 0.0, 0.0, float(j)] for j in range(18)]
    sub = np.asarray(subset, dtype=np.float64)
    if sub.size == 0:
        return [[0.0, 0.0, 0.0, float(j)] for j in range(18)]
    person = sub[0]
    out: list[list[float]] = []
    for j in range(18):
        idx = int(person[j])
        if idx < 0:
            out.append([0.0, 0.0, 0.0, float(j)])
        else:
            x, y, conf = float(cand[idx, 0]), float(cand[idx, 1]), float(cand[idx, 2])
            out.append([x, y, conf, float(j)])
    return out


def _load_train_captions(train_json: Path) -> dict[str, str]:
    """

    Load EquiFashion_DB/train.json captions.

    Returns: mapping {gt_filename -> caption}.

    """
    if not train_json.is_file():
        return {}
    data = json.loads(train_json.read_text(encoding="utf-8"))
    if not isinstance(data, list):
        return {}
    out: dict[str, str] = {}
    for r in data:
        if not isinstance(r, dict):
            continue
        gt = str(r.get("gt", "")).strip()
        if not gt:
            continue
        out[Path(gt).name] = str(r.get("caption", "")).strip()
    return out


def run_equifashion_train_pose(

    *,

    equi_root: Path,

    body_model: Path,

    hand_model: Path,

    write_summary_json: bool,

    overwrite_existing: bool,

    out_size: int,

    limit: int,

    only_gt: Optional[str],

    fix_bright: bool,

    bright_threshold: float,

) -> None:
    """

    Extract pose canvas (black background) + keypoints JSON for EquiFashion_DB/train.

    """
    train_dir = equi_root / "train"
    train_json = equi_root / "train.json"
    out_pose_dir = equi_root / "train_pose" / "pose"
    out_json_dir = equi_root / "train_pose" / "json"

    if not train_dir.is_dir():
        raise SystemExit(f"Không tìm thấy thư mục train: {train_dir}")
    if not body_model.is_file() or not hand_model.is_file():
        raise SystemExit(
            f"Cần file model OpenPose tại:\n  {body_model}\n  {hand_model}\n"
            "(tải body_pose_model.pth / hand_pose_model.pth từ pytorch-openpose / ControlNet bundle)."
        )

    cap_map = _load_train_captions(train_json)

    out_pose_dir.mkdir(parents=True, exist_ok=True)
    out_json_dir.mkdir(parents=True, exist_ok=True)

    op = OpenposeDetector(body_model_path=str(body_model), hand_model_path=str(hand_model))

    exts = {".jpg", ".jpeg", ".png", ".webp"}
    img_paths = [p for p in sorted(train_dir.iterdir()) if p.is_file() and p.suffix.lower() in exts]
    if only_gt:
        want = Path(only_gt).name
        img_paths = [p for p in img_paths if p.name == want]
    if limit > 0:
        img_paths = img_paths[:limit]
    if not img_paths:
        raise SystemExit(f"Không có ảnh trong {train_dir}")

    processed = 0
    skipped = 0
    failed = 0
    fixed = 0

    print(f"Found {len(img_paths)} image(s) to process. overwrite_existing={overwrite_existing}, size={out_size}", flush=True)

    for p in img_paths:
        gt = p.name
        stem = p.stem
        pose_path = out_pose_dir / gt
        json_path = out_json_dir / f"{stem}.json"
        # Optional repair: if an existing pose looks like the original image (too bright),
        # regenerate it (and JSON) even when overwrite is off.
        if fix_bright and pose_path.is_file():
            try:
                existing = cv2.imread(str(pose_path), cv2.IMREAD_COLOR)
                if existing is not None and float(existing.mean()) > float(bright_threshold):
                    # treat as wrong pose (likely original image), regenerate below
                    fixed += 1
                else:
                    # looks fine; if json exists too, we can skip safely
                    if (not overwrite_existing) and json_path.is_file():
                        skipped += 1
                        continue
            except Exception:
                pass

        # Skip only when both outputs exist and overwrite is off.
        if (not overwrite_existing) and pose_path.is_file() and json_path.is_file():
            skipped += 1
            continue

        bgr = cv2.imread(str(p), cv2.IMREAD_COLOR)
        if bgr is None:
            failed += 1
            continue

        # Resize input before OpenPose so keypoints match the saved canvas size.
        if out_size > 0 and (bgr.shape[0] != out_size or bgr.shape[1] != out_size):
            bgr = cv2.resize(bgr, (out_size, out_size), interpolation=cv2.INTER_AREA)

        # Black canvas output; keep models loaded for batch speed.
        canvas, keypoints = op(bgr, draw_on_image=False, auto_offload=False)

        # OpenPose returns RGB; OpenCV writes BGR.
        ok_pose = cv2.imwrite(str(pose_path), canvas[:, :, ::-1])
        if not ok_pose:
            failed += 1
            continue

        equif = openpose_to_equifashion_candidate(keypoints.get("candidate") or [], keypoints.get("subset") or [])
        json_path.write_text(json.dumps({"candidate": equif}, ensure_ascii=False), encoding="utf-8")
        processed += 1
        if processed % 200 == 0:
            print(f"progress: processed={processed}, fixed={fixed}, skipped={skipped}, failed={failed}", flush=True)

    if write_summary_json:
        summary: list[dict[str, str]] = []
        for p in img_paths:
            gt = p.name
            if not (out_pose_dir / gt).is_file():
                continue
            summary.append(
                {
                    "gt": gt,
                    "caption": cap_map.get(gt, ""),
                    "pose": f"train_pose/pose/{gt}".replace("\\", "/"),
                }
            )
        (equi_root / "train_pose.json").write_text(
            json.dumps(summary, ensure_ascii=False, indent=2),
            encoding="utf-8",
        )

    print(
        "EquiFashion train pose done: "
        f"processed={processed}, fixed={fixed}, skipped={skipped}, failed={failed}, out={equi_root / 'train_pose'}"
    )

    # Free memory at the end.
    try:
        op.offload()
    except Exception:
        pass


def main() -> None:
    ap = argparse.ArgumentParser(description="Batch OpenPose canvas for EquiFashion_DB/train -> train_pose")
    ap.add_argument(
        "--body-model",
        type=Path,
        default=ROOT / "model" / "body_pose_model.pth",
        help="OpenPose body weights",
    )
    ap.add_argument(
        "--hand-model",
        type=Path,
        default=ROOT / "model" / "hand_pose_model.pth",
        help="OpenPose hand weights",
    )

    ap.add_argument(
        "--equifashion-root",
        type=Path,
        default=ROOT / "EquiFashion_DB",
        help="EquiFashion_DB root (contains train/, train.json, ...).",
    )
    ap.add_argument(
        "--write-train-pose-json",
        action="store_true",
        help="Write EquiFashion_DB/train_pose.json (list of {gt, caption, pose}).",
    )
    ap.add_argument(
        "--overwrite-existing",
        action="store_true",
        help="Overwrite existing pose/json (default: skip).",
    )
    ap.add_argument(
        "--pose-size",
        type=int,
        default=512,
        help="Force input and output to size×size (default: 512).",
    )
    ap.add_argument(
        "--limit",
        type=int,
        default=0,
        help="Process only the first N images (0 = all).",
    )
    ap.add_argument(
        "--only-gt",
        type=str,
        default=None,
        help="Process only one file by gt name (e.g. 000611.jpg).",
    )
    ap.add_argument(
        "--fix-bright",
        action="store_true",
        help="Repair existing pose files that look too bright (likely the original image).",
    )
    ap.add_argument(
        "--bright-threshold",
        type=float,
        default=30.0,
        help="Mean-pixel threshold to treat a pose as 'bright' (default: 30).",
    )
    args = ap.parse_args()

    if not args.body_model.is_file() or not args.hand_model.is_file():
        raise SystemExit(
            f"Cần file model OpenPose tại:\n  {args.body_model}\n  {args.hand_model}\n"
            "(tải body_pose_model.pth / hand_pose_model.pth từ pytorch-openpose / ControlNet bundle)."
        )

    run_equifashion_train_pose(
        equi_root=Path(args.equifashion_root).resolve(),
        body_model=Path(args.body_model).resolve(),
        hand_model=Path(args.hand_model).resolve(),
        write_summary_json=bool(args.write_train_pose_json),
        overwrite_existing=bool(args.overwrite_existing),
        out_size=int(args.pose_size),
        limit=int(args.limit),
        only_gt=str(args.only_gt) if args.only_gt else None,
        fix_bright=bool(args.fix_bright),
        bright_threshold=float(args.bright_threshold),
    )


if __name__ == "__main__":
    main()