File size: 3,715 Bytes
33569f9 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 | import numpy as np
import pandas as pd
from libs.utils.Evaluation import average_recall_vs_avg_nr_proposals, compute_average_precision_detection
def compute_AP_AR(pred_time, gt_time, score,
iou_thresholds_ap=np.linspace(0.5, 0.95, 10),
iou_thresholds_ar=np.linspace(0.5, 0.95, 10),
ar_points=(10, 20, 50, 100),
subset='', max_avg_nr_proposals=100):
"""
完全复现 ActivityNet 官方 evaluation_detection 与 evaluation_proposal 的逻辑。
支持自定义 AR 评估点 ar_points。
返回所有结果为 0~1 之间的浮点数（不乘 100）。
"""
# ===== 构造 prediction DataFrame =====
video_ids, t_start, t_end, scores = [], [], [], []
for vid, (p, s) in enumerate(zip(pred_time, score)):
if len(p) == 0:
continue
for seg, sc in zip(p, s):
video_ids.append(f"vid{vid}")
t_start.append(float(seg[0]))
t_end.append(float(seg[1]))
scores.append(float(sc))
pred_df = pd.DataFrame({
'video-id': video_ids,
't-start': t_start,
't-end': t_end,
'score': scores,
'label': [0] * len(video_ids)
})
# ===== 构造 ground truth DataFrame =====
video_ids, t_start, t_end = [], [], []
for vid, g in enumerate(gt_time):
if len(g) == 0:
continue
for seg in g:
video_ids.append(f"vid{vid}")
t_start.append(float(seg[0]))
t_end.append(float(seg[1]))
gt_df = pd.DataFrame({
'video-id': video_ids,
't-start': t_start,
't-end': t_end,
'label': [0] * len(video_ids)
})
# ===== Detection (AP) =====
ap = compute_average_precision_detection(gt_df, pred_df, iou_thresholds_ap)
mAP = np.mean(ap)
mAP_at_tIoU = [f"mAP@{t:.2f} {v:.4f}" for t, v in zip(iou_thresholds_ap, ap)]
det_result = f"Detection ({subset}): average-mAP {mAP:.4f} {' '.join(mAP_at_tIoU)}"
# print(det_result)
# ===== Proposal (AR) =====
recall, avg_recall, proposals_per_video = average_recall_vs_avg_nr_proposals(
gt_df, pred_df,
max_avg_nr_proposals=max_avg_nr_proposals,
tiou_thresholds=iou_thresholds_ar
)
# 计算 AR@N，对应 recall 的固定索引（与官方一致）
AR_dict = {}
for n in ar_points:
idx = int(n - 1) # 因为 recall 的索引从 0 开始
AR_dict[f"AR@{n}"] = float(np.mean(recall[:, idx]))
AR_mean = np.mean(list(AR_dict.values()))
# 输出与官方格式一致
ar_str = " ".join([f"{k} {v:.4f}" for k, v in AR_dict.items()])
prop_result = f"Proposal ({subset}): {ar_str}"
# print(prop_result)
# ===== 汇总结果 =====
results = {
"mAP": float(mAP),
**{f"mAP@{str(round(t, 2)).rstrip('0').rstrip('.')}": float(v) for t, v in zip(iou_thresholds_ap, ap)},
**AR_dict,
"mAR": float(AR_mean)
}
return results
if __name__ == '__main__':
IOU_THRESHOLDS = np.linspace(0.5, 0.95, 10)
pred_time = [
np.array([[0.1, 0.4], [0.85, 1.0]]),
np.array([[0.2, 0.5], [0.6, 0.7], [0.9, 1.0]]),
np.array([[0.05, 0.15]])
]
gt_time = [
np.array([[0.15, 0.45], [0.5, 0.8]]),
np.array([[0.25, 0.55]]),
np.array([[0.08, 0.12]])
]
score = [
np.array([0.95, 0.60]),
np.array([0.85, 0.55, 0.40]),
np.array([0.70])
]
out = compute_AP_AR(
pred_time, gt_time, score,
iou_thresholds_ap=IOU_THRESHOLDS,
iou_thresholds_ar=IOU_THRESHOLDS,
ar_points=(1, 5, 10, 100),
subset='validation'
)
print(out)
|