9dff19e6ac
用户原话："赤裸的大人都是爸爸，家里没有其他人会赤裸"——加进闭集识别的免费 规则层（跟汤圆/媳妇同级），命中就不用再调用视觉大模型比对参考图，比 auto_id 更快更准。只对成年人生效（判断顺序上幼儿/儿童年龄档在前面，小孩光膀子玩很 正常，不适用这条规则）。 排查历史数据时用这条规则筛出了一个真实误判：video 1092 的一次赤膊出现被 auto_id（NVIDIA 视觉比对）错误识别成"爷爷"，实际是"爸爸"。已用 correct_video_identity 手动纠正（标记 source=manual，受保护不会被以后的自动 识别覆盖回去），描述文字已确认同步更新。 顺带确认了另一批命中同样关键词的历史数据（74 条里有 3 条是"汤圆"光膀子） 是正确的——3 岁小孩光膀子玩正常，是通过幼儿/儿童年龄档规则识别出来的，跟 赤膊无关，不是误判，没有动。 新增 4 个测试：赤膊成年男性直判爸爸且不调用大模型（用会报错的假 person_identifier 验证短路生效）、覆盖各种赤膊近义词表述、赤膊小孩仍然 正确判成汤圆不受影响、没有赤膊关键词时行为不变照常走大模型比对。
122 lines
4.8 KiB
Python
122 lines
4.8 KiB
Python
from datetime import datetime
|
||
|
||
from fam_edge.video_processor import (
|
||
VideoProcessor,
|
||
_parse_event_start_from_filename,
|
||
_parse_event_ts,
|
||
_clean_person,
|
||
)
|
||
|
||
|
||
class _RaisingPersonIdentifier:
|
||
"""用于验证"命中免费规则就不该再调用大模型比对"——一旦被调用直接报错，
|
||
测试能立刻发现规则短路失败。"""
|
||
def classify_adult_male(self, crop):
|
||
raise AssertionError("命中了免费规则的 uid 不该再走 person_identifier")
|
||
|
||
|
||
class _FixedPersonIdentifier:
|
||
def __init__(self, name):
|
||
self._name = name
|
||
|
||
def classify_adult_male(self, crop):
|
||
return self._name
|
||
|
||
|
||
def _bare_processor(person_identifier):
|
||
"""跳过 __init__（不需要真的加载 config/建适配器），只测
|
||
_resolve_closed_set_identities 这一个纯逻辑方法。"""
|
||
vp = VideoProcessor.__new__(VideoProcessor)
|
||
vp.person_identifier = person_identifier
|
||
return vp
|
||
|
||
|
||
def test_parse_filename_pure_digit_format():
|
||
assert _parse_event_start_from_filename(
|
||
"Generic_ONVIF-001-20260820-140416-1787205856321-7.mp4"
|
||
) == "2026-08-20 14:04:16"
|
||
|
||
|
||
def test_parse_filename_underscore_date_format():
|
||
assert _parse_event_start_from_filename("20260821_081500.mp4") == "2026-08-21 08:15:00"
|
||
|
||
|
||
def test_parse_filename_dashed_date_format():
|
||
assert _parse_event_start_from_filename("2026-08-21_081500.mp4") == "2026-08-21 08:15:00"
|
||
|
||
|
||
def test_parse_filename_no_match_returns_empty():
|
||
assert _parse_event_start_from_filename("客厅.mp4") == ""
|
||
|
||
|
||
def test_parse_event_ts_relative_offset():
|
||
start = datetime(2026, 8, 21, 15, 53, 3)
|
||
abs_ts, offset = _parse_event_ts("00:18:22", start)
|
||
assert abs_ts == "2026-08-21 16:11:25"
|
||
assert offset == 18 * 60 + 22
|
||
|
||
|
||
def test_parse_event_ts_relative_offset_no_start():
|
||
abs_ts, offset = _parse_event_ts("00:01:23", None)
|
||
assert abs_ts == "00:01:23"
|
||
assert offset == 83.0
|
||
|
||
|
||
def test_parse_event_ts_over_6_hours_falls_back_to_absolute():
|
||
"""相对时间 > 6 小时视为模型误输出绝对时间，不强行按偏移定位。"""
|
||
start = datetime(2026, 8, 21, 8, 0, 0)
|
||
abs_ts, offset = _parse_event_ts("2026-08-21 09:00:00", start)
|
||
assert abs_ts == "2026-08-21 09:00:00"
|
||
assert offset == 3600.0
|
||
|
||
|
||
def test_clean_person_strips_fullwidth_parens():
|
||
assert _clean_person("人物A（别名/标识：人物B）") == "人物A"
|
||
|
||
|
||
def test_clean_person_strips_ascii_parens():
|
||
assert _clean_person("人物A(alias: 人物B)") == "人物A"
|
||
|
||
|
||
def test_clean_person_no_parens_unchanged():
|
||
assert _clean_person("汤圆") == "汤圆"
|
||
|
||
|
||
# ----------------------------------------------------------------------
|
||
# 闭集人物识别：赤膊成年人规则（用户原话："赤裸的大人都是爸爸，家里没有
|
||
# 其他人会赤裸"）——命中即免费直判，不调用视觉大模型比对。
|
||
# ----------------------------------------------------------------------
|
||
|
||
def test_shirtless_adult_male_resolves_to_dad_without_vlm_call():
|
||
vp = _bare_processor(_RaisingPersonIdentifier())
|
||
resolved = vp._resolve_closed_set_identities(
|
||
1, {"人物A": {"gender": "男", "age_band": "中年", "clothing": "赤膊+深色长裤"}}, [])
|
||
assert resolved == {"人物A": ("爸爸", "rule")}
|
||
|
||
|
||
def test_shirtless_variants_all_match():
|
||
for phrase in ("光着上身", "赤膊", "裸体", "光膀子", "上身赤裸", "未穿上衣", "深色长裤，赤裸上身"):
|
||
vp = _bare_processor(_RaisingPersonIdentifier())
|
||
resolved = vp._resolve_closed_set_identities(
|
||
1, {"人物A": {"gender": "男", "age_band": "中年", "clothing": phrase}}, [])
|
||
assert resolved.get("人物A") == ("爸爸", "rule"), f"未命中: {phrase}"
|
||
|
||
|
||
def test_shirtless_child_still_resolves_to_child_not_dad():
|
||
"""核心诉求: 用户的规则明确是"赤裸的大人"，小孩光膀子玩很正常，不适用
|
||
这条规则——幼儿/儿童年龄档要走在赤膊判断前面，不能被误判成爸爸。"""
|
||
vp = _bare_processor(_RaisingPersonIdentifier())
|
||
resolved = vp._resolve_closed_set_identities(
|
||
1, {"人物A": {"gender": "男", "age_band": "幼儿", "clothing": "赤裸上身+深色短裤"}}, [])
|
||
assert resolved == {"人物A": ("汤圆", "rule")}
|
||
|
||
|
||
def test_clothed_adult_male_still_falls_through_to_vlm(monkeypatch):
|
||
"""核心诉求: 没有赤膊关键词的正常穿戴场景，行为不变——照常走视觉大模型
|
||
比对（这里用假的 _best_crop_for_uid 避免真的需要 norm_events 数据）。"""
|
||
vp = _bare_processor(_FixedPersonIdentifier("爷爷"))
|
||
monkeypatch.setattr(vp, "_best_crop_for_uid", lambda *a, **k: b"fake-jpeg-bytes")
|
||
resolved = vp._resolve_closed_set_identities(
|
||
1, {"人物A": {"gender": "男", "age_band": "中年", "clothing": "蓝色Polo衫"}}, [])
|
||
assert resolved == {"人物A": ("爷爷", "auto_id")}
|