📦 模型下載：https://pub-478929a98a5c440cb22c2241c0bde314.r2.dev/safety_rope_v20260519/best.pt

🪢 Safety Rope v20260519 (DINOv3-S + RoI HD)

訓練日期 2026-05-20 · 5090-2 GPU 0 · backbone vit_small_patch16_dinov3 22.47M

🚀 結論：在更大 test set 上仍進步 +1.5pp AP，FP -13.7%

cvat #10 新通過 acceptance 的 11 個 SIEMENS_UNKNOWN_CH01-04 task / 1905 frame 進入訓練（+1029 train / +305 val / +571 test frame）。即使 test set 變大 (5239→5637) test_AP 仍從 v518 的 0.9024 提升到 0.9170，Precision 顯著改善 +3.6pp (0.814 → 0.850)，FP 降 54 (395 → 341)。

📊 核心指標（test set）

0.9170
test_AP
0.850
Precision
0.863
Recall
0.856
F1
0.885
Accuracy

🆚 v518 vs v519 對比

指標v518v519Δ
best_val_AP0.9316 (ep03)0.9474 (ep07)+1.6pp
test_AP0.90240.9170+1.5pp
test Precision0.8140.850+3.6pp
test Recall0.8810.863-1.8pp
test F10.8460.856+1.0pp
test FP（誤報）395341-54 (-13.7%)
test FN（漏報）233306+73
test n_total52395637+398 (+7.6%)
epochs_run11 (early stop)15 (early stop)

📂 Dataset（cvat #10 + cvat #8）

Splitv518 framev519 frame變化
train~15,16416,193+1,029 SIEMENS
val~4,2564,561+305 SIEMENS
test5,2395,637+571 SIEMENS

新增 11 個 SIEMENS_UNKNOWN_CH01-04 task

tidsubsetsizechannel
2861Train400CH01
2862Test111CH01
2863Validation66CH01
2864Train327CH02
2865Test44CH02
2866Validation42CH02
2867Test190CH03
2868Validation132CH03
2869Train302CH04
2870Test226CH04
2871Validation65CH04

🔍 解讀

📦 模型下載

https://pub-478929a98a5c440cb22c2241c0bde314.r2.dev/safety_rope_v20260519/best.pt ⬇

⚙️ Hyperparams（完全沿用 v518 baseline）

backbone: vit_small_patch16_dinov3 (22.47M params)
img_size: 1280×720 (HD)
expand: x=1.0, y_top=0.2, y_bot=1.5
jitter: center=0.2, size=[0.7,1.4]
batch: 8, epochs: 30, patience: 8
lr: 1e-4, wd: 0.01
photometric: brightness ±0.4, contrast ±0.3, saturation ±0.4

best_val_AP @ ep07 = 0.9474
test thr = 0.429 (best F1)

← 訓練報告目錄 · 前版 v20260518 · 訓練 SOP