👷 PPE 22-attr v20260519
MobileNetV3-L + 22 binary heads + cvat #12 5/19 標記師補強
結果摘要:best ep10 test_mAP=0.9505 / macro_F1=0.9214 — aluminized_apron 解鎖、cotton_gloves 改善
對標 baseline:v20260515b baseline val_mAP=0.981
📊 結果
| 指標 | v20260515b baseline | v20260519 best (ep10) |
| val_mAP | 0.981 | 0.9503 |
| test_mAP | ~0.961 | 0.9505 |
| macro F1 | 0.921 | 0.9214 |
| aluminized_apron AP | 0 (不可評估) | 1.000 ⭐ |
| cotton_gloves AP | R@P95=0.13 | 0.860 |
| helmet_goggles AP | — | 0.986 (樣本 15→2346) |
| hair_cover AP | — | 0.921 (樣本 35→184) |
| sleeves AP | — | 0.712 (仍弱) |
📦 訓練 stack
- Backbone:MobileNetV3-L (4.23M params)
- Stack:Multi-head 22 binary + partial-label BCE + camaug + mixup + per-attr neg-weight
- Hyperparams:40 ep / patience 8 / batch=128 / imgsz=[384,192] / AdamW lr=3e-4 wd=0.01 / mixup=0.2
- Neg-weight 對齊 v504:harness=2.0 / hard_hat=1.3 / safety_vest=1.3
- Dataset:cvat #12 + 外部 5 source: 99,812 train / 11,580 val / 12,514 test
- 5/19 標記師補的 attr:aluminized_apron 0→1012 / sleeves 67→520 / helmet_goggles 15→2346 / hair_cover 35→184
- 訓練時間:20 min (early stop ep18/40)
📝 觀察
核心收益:aluminized_apron 從不可評估 → ceiling、cotton_gloves R@P95 0.13 → AP 0.86、helmet_goggles / hair_cover / sleeves / rubber_gloves 補資料後可信。test_mAP 略低於 v515b 是因 test set 變難(含新 attr)。