fall/data/pose_extract.py

from __future__ import annotations

from pathlib import Path

import numpy as np
import pandas as pd
from tqdm import tqdm

from fall.data.video import read_sampled_frames


def extract_yolo_pose(frames: np.ndarray, model, device: str) -> np.ndarray:
    images = [frame for frame in frames]
    results = model.predict(images, device=device, verbose=False)
    all_kpts = []
    for res in results:
        if res.keypoints is None or res.keypoints.data is None or len(res.keypoints.data) == 0:
            k = 17
            all_kpts.append(np.zeros((k, 3), dtype=np.float32))
            continue
        data = res.keypoints.data.detach().cpu().numpy()
        scores = data[..., 2].mean(axis=1)
        person = data[int(scores.argmax())]
        all_kpts.append(person[:, :3].astype(np.float32))
    return np.stack(all_kpts, axis=0)


def extract_pose_manifest(
    manifest: str | Path,
    out_dir: str | Path,
    out_manifest: str | Path,
    frames: int = 32,
    backend: str = "yolo",
    model_name: str = "yolov8n-pose.pt",
    device: str = "cuda:0",
    resize: int | None = None,
) -> pd.DataFrame:
    if backend != "yolo":
        raise ValueError("Only backend='yolo' is implemented in this reproducible pipeline.")
    from ultralytics import YOLO

    model = YOLO(model_name)
    df = pd.read_csv(manifest)
    out_dir = Path(out_dir)
    out_dir.mkdir(parents=True, exist_ok=True)
    rows = []
    for i, row in tqdm(df.iterrows(), total=len(df), desc="extract pose"):
        video_id = str(row.get("video_id", Path(row["video_path"]).stem))
        pose_path = out_dir / f"{video_id}.npy"
        if pose_path.exists():
            pose = np.load(pose_path)
        else:
            start = row.get("start", None)
            end = row.get("end", None)
            fps = row.get("fps", None)
            start = None if pd.isna(start) else float(start)
            end = None if pd.isna(end) else float(end)
            fps = None if pd.isna(fps) else float(fps)
            sampled = read_sampled_frames(row["video_path"], target_frames=frames, resize=resize, start=start, end=end, fps=fps)
            pose = extract_yolo_pose(sampled, model=model, device=device)
            np.save(pose_path, pose.astype(np.float32))
        new_row = dict(row)
        new_row["pose_path"] = str(pose_path)
        new_row["pose_shape"] = "x".join(map(str, pose.shape))
        rows.append(new_row)
    out_df = pd.DataFrame(rows)
    out_manifest = Path(out_manifest)
    out_manifest.parent.mkdir(parents=True, exist_ok=True)
    out_df.to_csv(out_manifest, index=False)
    return out_df