🦺 Factory PPE 21-attr — v20260501 (+fall 跌倒) 報告

partial-label BCE / 21 binary heads / per-attr mask(unknown→mask=0 不算 loss/metric)

v20260501 主改動:在 cvat2 project 5 person label 加入 fall(跌倒判定)attribute,並從 raicvat#6 fall dataset(48 tasks 含 MRT/機場/GUANXI 監視器), 並用 v20260415 ppe_attr(gx10 訓練的舊版 4-cls model)對所有 raicvat_* 76,184 person crops inference 取得 harness probability, threshold 0.5 → 寫回 cvat2(PATCH action=update)。 其他 PPE source(既有 133k shapes)fall=unknown(mask=0 不算 loss)(partial-label 機制讓不相關樣本不影響 fall head 學習)。

資料:cvat2 project 5(30 tasks, 9 sources, ~163k crops)+ fall_p6 31,797 shapes (yes 2,318 / no 29,479)。

🏆 推薦 deploy

val_mAP (20 attr)
0.9663
test_mAP
0.9775
macro_f1
0.9535
harness test_AP
0.947
params
4.2M
train time
47min

v20260430_mobilenetv3l — 12 版本中 cost-effective 最佳,含完整 21 個 PPE/姿態 attribute 識別能力(含 harness + fall)。 checkpoint: 5090-2:~/factory_ppe/runs/factory_ppe_v20260430_mobilenetv3l/best.pt

模型下載 (R2 public)

模型架構

12 版本完整對照

versionbackbonedatan_attrval_mAPtest_mAPmacro_f1paramsepochstime
v20260426_mobilenetv3lMobileNetV3-Lbase 6src190.66550.98500.95324.2M21 (best=13)22min
v20260426_efficientnetb0EfficientNet-B0base 6src190.65120.98300.96114.0M33 (best=25)37min
v20260426_convnexttinyConvNeXt-Tinybase 6src190.69740.98800.960527.9M22 (best=16)68min
v20260427_efficientnetb0_cppe5EfficientNet-B0+CPPE-5190.76400.98290.95374.0M17 (best=9)19min
v20260427_convnexttiny_cppe5ConvNeXt-Tiny+CPPE-5190.76020.98050.942127.9M11 (best=5)35min
v20260427_mobilenetv3l_cppe5MobileNetV3-L+CPPE-5190.77290.97880.94804.2M18 (best=10)19min
v20260428_convnexttinyConvNeXt-Tiny+R2PPE190.92400.99250.977327.9M16 (best=10)61min
v20260428_mobilenetv3lMobileNetV3-L+R2PPE190.92380.99450.97704.2M17 (best=9)22min
v20260429_convnexttinyConvNeXt-Tiny+4-task split190.97020.97790.961527.9M24 (best=18)86min
v20260429_mobilenetv3lMobileNetV3-L+4-task split190.97120.98330.96504.2M27 (best=19)33min
v20260430_convnexttinyConvNeXt-Tiny+harness200.96480.97420.946227.9M12 (best=6)43min
v20260430_mobilenetv3lMobileNetV3-L+harness200.97010.97850.95414.2M32 (best=24)39min
v20260501_mobilenetv3l ⭐MobileNetV3-L+fall210.96630.97750.95354.2M32 (best=24)47min

訓練曲線(12 版本)

12-version val_mAP curves

v20260430 vs v20260501 (fall 整合的影響)

final epoch val_AP per attribute:

attributev430 (20-attr)v501 (21-attr)delta
hard_hat
安全帽
0.99330.9933+0.0000
no_head_protection
無護頭
0.99680.9968+0.0000
full_face_mask
全面罩
0.99830.9983+0.0000
face_mask
口罩
0.99170.9917+0.0000
no_gloves
無手套
0.99100.9910+0.0000
cotton_gloves
棉手套
0.81510.8151+0.0000
rubber_gloves
橡膠手套
1.00001.0000+0.0000
no_protective_clothing
無防護衣
0.99990.9999+0.0000
cleanroom_suit
無塵衣
1.00001.0000+0.0000
splash_proof_gown
防潑罩袍
1.00001.0000+0.0000
safety_vest
反光背心
0.97590.9759+0.0000
safety_shoes
安全鞋
0.82330.8233+0.0000
no_safety_shoes
無安全鞋
0.99000.9900+0.0000
no_sleeves
無絕緣袖
0.99970.9997+0.0000
heartbeat
生命徵象器
0.96690.9669+0.0000
sleeves
絕緣袖
1.00001.0000+0.0000
safety_glasses
護目鏡
0.84950.8495+0.0000
hair_cover
髮帽
1.00001.0000+0.0000
helmet_goggles
頭盔護目
1.00001.0000+0.0000
harness
安全帶
0.93230.9323+0.0000
fall
跌倒
0.92780.9278+0.0000

v20260430 m3l Test Per-attribute Metrics

test split 跨 8 source ~13k samples:

attributeAPF1PRthrvalid
hard_hat
安全帽
0.9970.9790.9790.9800.536101
no_head_protection
無護頭
0.9960.9740.9770.9720.546101
full_face_mask
全面罩
1.0000.9930.9950.9920.832626
face_mask
口罩
0.9980.9800.9740.9860.401604
no_gloves
無手套
0.9990.9860.9780.9940.452894
cotton_gloves
棉手套
0.8860.7910.8500.7390.851813
rubber_gloves
橡膠手套
1.0000.9980.9951.0000.482431
no_protective_clothing
無防護衣
0.9990.9960.9980.9950.403405
cleanroom_suit
無塵衣
1.0000.9971.0000.9940.861917
splash_proof_gown
防潑罩袍
1.0001.0001.0001.0000.852322
safety_vest
反光背心
0.9790.9470.9460.9480.545784
safety_shoes
安全鞋
0.8340.7840.7020.8890.921813
no_safety_shoes
無安全鞋
1.0000.9980.9960.9990.261813
no_sleeves
無絕緣袖
1.0000.9930.9910.9950.461813
heartbeat
生命徵象器
0.9390.8800.8370.9280.872662
sleeves
絕緣袖
0.9670.9090.8331.0000.031813
safety_glasses
護目鏡
0.9970.9790.9760.9810.641500
hair_cover
髮帽
1.0001.0001.0001.0001.001813
helmet_goggles
頭盔護目
1.0000.9961.0000.9920.872073
harness
安全帶
0.9470.8790.8900.8680.8113188
fall
跌倒
0.9900.9640.9530.9760.872097

0/20 attribute 在 test 仍無 mask=1 sample

harness 整合說明

下一步建議

報告生成:2026-04-27|v20260429v20260428v20260427label guide