feat: 脸型分类器(7分类)+批量报告生成

基于 MediaPipe 468 关键点提取几何特征,转 z 分数后与各脸型原型加权匹配。
参考分布与原型靶心取自 1093 张测试集的实测画像,不再靠人工设定绝对阈值。

方形脸占比从 29.6% 降到 13.8%,两处原因:一是参考统计量原先只由 50 张样本
估得,相对全量人群有系统性偏移,且三项偏移都在给方形脸加分;二是原型把
aspect_ratio 当作方形脸的主特征,但实测方脸组该值中位仅 +0.16,真正"宽"的
是圆脸(+1.01),等于在拿脸宽找方脸。

原型参数在「6 张基准标注图判定不变、且领先第二名 >=3 分」的约束下搜索得到。
余量约束是必要的:早前一版余量仅 0.008 分,权重写码时四舍五入就会翻转结论。

测试素材(人像照片)与报告输出体积大,一并加入 .gitignore。

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
xsl
2026-07-29 13:16:43 +08:00
co-authored by Cursor
parent 81c7a0b55c
commit 587876f8a0
12 changed files with 2755 additions and 0 deletions
+83
View File
@@ -0,0 +1,83 @@
"""
把各数据集的人脸特征抽取一次并缓存为 JSON,供调参脚本反复使用。
MediaPipe 关键点检测是调参循环里唯一的耗时环节,缓存后调参可以秒级迭代。
"""
from __future__ import annotations
import json
import sys
from pathlib import Path
from typing import Dict, List
import cv2
sys.path.insert(0, str(Path(__file__).resolve().parent))
from face_shape_classifier import ( # noqa: E402
_get_face_mesh,
extract_face_features,
)
ROOT = Path(__file__).resolve().parent
IMG_EXT = {".png", ".jpg", ".jpeg", ".webp", ".bmp"}
def iter_images(root: Path) -> List[Path]:
return sorted(p for p in root.rglob("*") if p.suffix.lower() in IMG_EXT)
def features_for(path: Path) -> Dict[str, float] | None:
bgr = cv2.imread(str(path))
if bgr is None:
return None
h, w = bgr.shape[:2]
rgb = cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)
res = _get_face_mesh().process(rgb)
if not res.multi_face_landmarks:
return None
return extract_face_features(res.multi_face_landmarks[0].landmark, image_size=(w, h))
def main() -> None:
out_path = Path(sys.argv[1]) if len(sys.argv) > 1 else ROOT / "cache" / "features.json"
out_path.parent.mkdir(parents=True, exist_ok=True)
sources = {
# 6 张带标注的基准图,文件名即期望脸型
"benchmark": [p for p in ROOT.joinpath("test_img").glob("*.png")],
"girl": iter_images(ROOT / "test_img" / "girl"),
"man": iter_images(ROOT / "test_img" / "man"),
"dataset": iter_images(ROOT / "test_img" / "脸型测试集合"),
}
records = []
failed = 0
for source, paths in sources.items():
for i, path in enumerate(paths, 1):
feats = features_for(path)
if feats is None:
failed += 1
continue
rel = path.relative_to(ROOT)
records.append(
{
"source": source,
"path": rel.as_posix(),
"file": path.name,
# dataset 的上级目录名即原始分组(弱标签,非可信真值)
"group": path.parent.name if source == "dataset" else source,
"expected": path.stem if source == "benchmark" else None,
"features": feats,
}
)
if i % 50 == 0 or i == len(paths):
print(f"[{source}] {i}/{len(paths)}", flush=True)
out_path.write_text(json.dumps(records, ensure_ascii=False), encoding="utf-8")
print(f"\n写入 {out_path}{len(records)} 条,检测失败 {failed}")
if __name__ == "__main__":
main()