Spaces:
Sleeping
Sleeping
File size: 3,700 Bytes
11fab85 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 | """Event-level 評估:預測事件與 GT 事件的配對與指標計算。
協定(README 同步聲明):
- 預測事件與「GT 事件 ± tol」有任何時間交集即為候選配對;
- 每個 GT 以最大交集貪婪配一個預測(一對一);
- TP = 配對成功的 GT 數;FN = 未配對 GT;
FP = fall 影片中未配對的預測 + ADL 影片中的所有預測
(同一 GT 的重複預測算 FP:懲罰事件碎裂);
- 另報 video-level specificity(無任何預測事件的 ADL 影片比例),對齊文獻報法。
"""
from __future__ import annotations
from dataclasses import dataclass, field
Interval = tuple[float, float]
def interval_overlap(a: Interval, b: Interval) -> float:
"""兩區間的交集長度(秒);不相交為 0。"""
lo = max(a[0], b[0])
hi = min(a[1], b[1])
return max(0.0, hi - lo)
@dataclass
class MatchResult:
tp: int
fp: int
fn: int
matches: list[tuple[Interval, Interval]] = field(default_factory=list) # (pred, gt)
def match_events(
preds: list[Interval], gts: list[Interval], tol_s: float = 0.5
) -> MatchResult:
"""單支影片內的貪婪一對一配對。"""
used = [False] * len(preds)
matches: list[tuple[Interval, Interval]] = []
for gt in sorted(gts):
expanded = (gt[0] - tol_s, gt[1] + tol_s)
best_i, best_ov = None, 0.0
for i, p in enumerate(preds):
if used[i]:
continue
ov = interval_overlap(p, expanded)
if ov > best_ov:
best_i, best_ov = i, ov
if best_i is not None and best_ov > 0.0:
used[best_i] = True
matches.append((preds[best_i], gt))
tp = len(matches)
fn = len(gts) - tp
fp = sum(1 for u in used if not u)
return MatchResult(tp=tp, fp=fp, fn=fn, matches=matches)
def evaluate_videos(videos: list[dict], tol_s: float = 0.5) -> dict:
"""彙整多支影片的 event-level 指標。
Args:
videos: 每支影片一個 dict:
``{"name": str, "is_adl": bool, "preds": [(s,e)...], "gts": [(s,e)...]}``
tol_s: GT 區間兩端的容忍(秒)。
Returns:
指標 dict(precision/recall/f1 在分母為 0 時為 None,不假造數字),
含 per-video 明細供失敗分析挑案例。
"""
tp = fp = fn = 0
adl_total = adl_clean = 0
per_video = []
for v in videos:
res = match_events(list(v.get("preds", [])), list(v.get("gts", [])), tol_s)
tp += res.tp
fp += res.fp
fn += res.fn
if v.get("is_adl", False):
adl_total += 1
if not v.get("preds"):
adl_clean += 1
per_video.append(
{
"name": v["name"],
"is_adl": bool(v.get("is_adl", False)),
"tp": res.tp,
"fp": res.fp,
"fn": res.fn,
"preds": [list(p) for p in v.get("preds", [])],
"gts": [list(g) for g in v.get("gts", [])],
}
)
precision = tp / (tp + fp) if (tp + fp) > 0 else None
recall = tp / (tp + fn) if (tp + fn) > 0 else None
f1 = (
2 * precision * recall / (precision + recall)
if precision is not None and recall is not None and (precision + recall) > 0
else None
)
specificity = adl_clean / adl_total if adl_total > 0 else None
return {
"tol_s": tol_s,
"tp": tp,
"fp": fp,
"fn": fn,
"precision": precision,
"recall": recall,
"f1": f1,
"video_level_specificity": specificity,
"n_videos": len(videos),
"n_adl_videos": adl_total,
"per_video": per_video,
}
|