#!/usr/bin/env python3 """Measure where the speaker is, and report where an overlay card may actually go. Guessing this is how a graphic ends up on someone's forehead. A "framed vertical talking head" is assumed to put the face in the upper third, but real footage varies enormously: on the material this was built against the subject occupied y 350..1653 of 1920, leaving one usable band of 350 px above the head and nothing else. The subject is located by skin-tone probability averaged over several frames, then the free bands above and below are reported together with the platform's safe area and the caption zone, so a card can be placed against numbers instead of a guess. Skin detection is deliberately crude — an RGB rule, not a face model. It does not need to find a face; it needs to find *where the person is*, and for that a coarse mask averaged over frames is both sufficient and dependency-free. """ from __future__ import annotations import argparse import json import subprocess import sys from pathlib import Path sys.path.insert(0, str(Path(__file__).resolve().parent)) from media_probe import probe # noqa: E402 from skill_config import canvas as resolve_canvas, emit, require_binary, utf8_stdout # noqa: E402 def sample_frame(video: Path, at: float, width: int, height: int): import numpy completed = subprocess.run( ["ffmpeg", "-v", "error", "-nostdin", "-ss", f"{at:.3f}", "-i", str(video), "-frames:v", "1", "-f", "rawvideo", "-pix_fmt", "rgb24", "-"], capture_output=True, ) expected = width * height * 3 if completed.returncode or len(completed.stdout) < expected: return None return numpy.frombuffer(completed.stdout[:expected], dtype=numpy.uint8) \ .reshape(height, width, 3).astype(numpy.int16) def skin_mask(frame): """Coarse RGB skin rule. Finds the person, not the face.""" import numpy red, green, blue = frame[:, :, 0], frame[:, :, 1], frame[:, :, 2] brightest = frame.max(axis=2) darkest = frame.min(axis=2) return ( (red > 70) & (green > 36) & (blue > 18) & ((brightest - darkest) > 12) & (numpy.abs(red - green) > 10) & (red > green) & (red > blue) ) def free_bands(row_probability, threshold: float, height: int) -> list[dict]: """Contiguous rows where the subject is essentially absent.""" import numpy occupied = row_probability > threshold bands = [] start = None for y in range(height): if not occupied[y] and start is None: start = y elif occupied[y] and start is not None: bands.append({"top": start, "bottom": y, "height": y - start}) start = None if start is not None: bands.append({"top": start, "bottom": height, "height": height - start}) return [band for band in bands if band["height"] >= 60] def main() -> None: utf8_stdout() parser = argparse.ArgumentParser(description=__doc__.splitlines()[0]) parser.add_argument("video") parser.add_argument("--canvas", default="reels") parser.add_argument("--samples", type=int, default=10) parser.add_argument("--threshold", type=float, default=0.06, help="Доля кадров со «кожей» в строке, выше которой строка занята") parser.add_argument("--skip", default="", help="Интервалы, которые не считать (там б-ролл, а не спикер): " "'4.5-8,17.8-21'") parser.add_argument("--caption-map", help="Учесть зону титров как занятую") parser.add_argument("--report") args = parser.parse_args() try: import numpy except ImportError: raise SystemExit("Нужен numpy: pip install numpy") require_binary("ffmpeg") video = Path(args.video).resolve() if not video.is_file(): raise SystemExit(f"Not found: {video}") info = probe(video) stream = info.get("video") or {} width = stream.get("display_width") height = stream.get("display_height") duration = float(info.get("duration") or 0.0) if not width or not height or duration <= 0: raise SystemExit("Не удалось прочитать геометрию или длительность") board = resolve_canvas(args.canvas) skip: list[tuple[float, float]] = [] for chunk in args.skip.split(","): if "-" in chunk: low, _, high = chunk.partition("-") try: skip.append((float(low), float(high))) except ValueError: continue times = [] step = duration / (args.samples + 1) for index in range(1, args.samples + 1): at = step * index if any(low <= at <= high for low, high in skip): continue times.append(round(at, 3)) if not times: raise SystemExit("Все выбранные моменты попали в --skip") accumulator = numpy.zeros((height, width), dtype=numpy.float32) used = 0 for at in times: frame = sample_frame(video, at, width, height) if frame is None: continue accumulator += skin_mask(frame) used += 1 if used == 0: raise SystemExit("Не удалось декодировать ни один кадр") probability = accumulator / used rows = probability.mean(axis=1) occupied = numpy.where(rows > args.threshold)[0] subject = { "top": int(occupied.min()) if occupied.size else None, "bottom": int(occupied.max()) if occupied.size else None, } bands = free_bands(rows, args.threshold, height) caption_zone = None if args.caption_map: payload = json.loads(Path(args.caption_map).read_text(encoding="utf-8")) ys = [word["y"] for block in payload.get("blocks", []) for word in block.get("words", []) if word.get("y")] if ys: size = payload.get("nominal_font_size", 60) caption_zone = {"top": int(min(ys) - size), "bottom": int(max(ys) + size)} safe = board.safe_box scored = [] for band in bands: top = band["top"] bottom = band["bottom"] # Trim against the caption zone: a card over the captions is as bad as a card # over the face. if caption_zone: if top < caption_zone["bottom"] and bottom > caption_zone["top"]: if caption_zone["top"] - top > bottom - caption_zone["bottom"]: bottom = min(bottom, caption_zone["top"]) else: top = max(top, caption_zone["bottom"]) usable = bottom - top if usable < 60: continue # The bottom of a vertical frame belongs to the platform's own interface. in_platform_ui = bottom > safe["bottom"] scored.append({ "top": top, "bottom": bottom, "height": usable, "overlaps_platform_ui": in_platform_ui, "recommended": usable >= 200 and not in_platform_ui, }) scored.sort(key=lambda item: (-int(item["recommended"]), -item["height"])) best = next((item for item in scored if item["recommended"]), None) advice: list[str] = [] if best: advice.append( f"Ставь карточку в y {best['top']}..{best['bottom']} и держи её высоту " f"не больше {best['height'] - 40} px с запасом на тень." ) else: advice.append( "Свободной полосы нужного размера нет: спикер занимает почти весь кадр. " "Варианты — сделать карточку узкой и вынести в угол, уменьшить её высоту " "под самую большую полосу, или отвести карточке отдельный кадр вместо " "наложения." ) if subject["top"] is not None and subject["top"] < 120: advice.append("Голова начинается почти от верхней кромки: сверху места нет.") report = { "video": str(video), "geometry": f"{width}x{height}", "frames_used": used, "sampled_at": times, "subject_rows": subject, "subject_height": (subject["bottom"] - subject["top"]) if subject["top"] is not None else None, "free_bands": scored, "best_band": best, "platform_safe_box": safe, "caption_zone": caption_zone, "row_profile": [ {"y": y, "occupancy": round(float(rows[y:y + 96].mean()), 3)} for y in range(0, height, 96) ], "advice": advice, "method": ( "Грубая маска кожи по RGB, усреднённая по кадрам. Задача не найти лицо, " "а найти, где находится человек — для этого достаточно и не нужны " "дополнительные зависимости." ), } if args.report: Path(args.report).parent.mkdir(parents=True, exist_ok=True) Path(args.report).write_text(json.dumps(report, ensure_ascii=False, indent=2) + "\n", encoding="utf-8") emit(report) if __name__ == "__main__": main()