📦 模型下載：https://pub-478929a98a5c440cb22c2241c0bde314.r2.dev/person_yolo26n_v20260521/best.pt

👤 Person YOLO26n v20260521

訓練日期 2026-05-21 · 5090-2 GPU 1 單卡 batch 64 · base = yolo26n.pt

⭐ 結論：當前最佳，Precision +0.6pp vs v518，mAP50-95 +0.5pp vs v520

cvat #1 自 v520 啟動後新通過 27 task / 1500 frame（SIEMENS_UNKNOWN_CH01-02 系列）全部納入訓練。fair compare（v521 test set）mAP50 0.8751 略勝 v518 0.8750 / v520 0.8709，Precision 0.919 顯著最佳對 ppe-demo cascade 降誤報有意義。

📊 核心指標（v521 test set, 5863 img / 21886 instances）

0.919

Precision

0.777

Recall

0.8751

mAP@0.50

0.6806

mAP@0.50-95

🆚 4-way 對比（v521 test set fair compare）

Ckpt	P	R	mAP50	mAP50-95	備註
v518 yolo11n	0.914	0.787	0.8750	0.6757	YOLO11n baseline
v519 yolo26n	0.908	0.780	0.8676	0.6790	YOLO26 +SAM3
v520 yolo26n	0.905	0.780	0.8709	0.6776	+10 SIEMENS CH01-02
v521 yolo26n	0.919	0.777	0.8751	0.6806	+27 task 完整 SIEMENS ⭐

v521 Precision +0.4pp vs v520 / +0.6pp vs v518。mAP50 跟 v518 持平，mAP50-95 最佳。Recall 略降但 P/R balance 對 ppe-demo cascade 後續 classification 更有利。

📂 Dataset

Split	v520	v521	Δ
train	26,505	27,558	+1,053
val	3,847	3,982	+135
test	5,815	5,863	+48

新增 17 個 SIEMENS_UNKNOWN_CH01-02 task（5/20-5/21 標完，v520 啟動後）

tid	subset	size	name
4781-4782	test	51	CH02_002, CH02_012
4793	train	49	CH01_006
4794-4798	train	289	CH02_003-009
4800	train	18	CH02_013
4803-4805	train	185	CH01_001-003
4806-4810	train	436	CH02_011, 015-018

📦 模型下載

person_yolo26n_v20260521/best.pt ⬇

⚙️ Hyperparams（完全沿用 v520 baseline，僅改 device 單卡）

task: detect, model: yolo26n.pt, epochs: 100, patience: 30
batch: 64, imgsz: 640, device: 1 (單卡 GPU 1, GPU 0 同時跑 fire_smoke v521)
cache: ram, workers: 8
optimizer: auto, lr0: 0.01, lrf: 0.01, momentum: 0.937, weight_decay: 0.0005
mosaic: 1.0, close_mosaic: 10, fliplr: 0.5, translate: 0.1, scale: 0.5
hsv_h: 0.015, hsv_s: 0.7, hsv_v: 0.4
iou: 0.7, max_det: 300, seed: 0

# SUBSET_MAP 已支援 lowercase + SAM3
{"Train": "train", "Validation": "val", "Test": "test", "SAM3": "train",
 "train": "train", "validation": "val", "test": "test"}

# 訓練時間 ~2.7 hr (100 ep 單卡 GPU 1)
# best mAP50 @ ep80 = 0.691 val

← 訓練報告目錄 · 前版 v20260520 · 訓練 SOP