🚪 Hatch 門狀態偵測 — image classifier 報告

架構:MobileNetV3-L backbone + 2 binary heads (image-level multi-label),跟 fire_smoke v20260410 同 pattern。

資料:cvat2 project 7「door_state_detection」7,471 frames(從 raicvat#12 hatch project 168 tasks 移植)。每張圖標 1 個 door tag with state attr ∈ [open/closed/mixed/unknown]。

🏆 結果

0.9996
val_mAP
0.9869
test_mAP
0.9482
macro_f1
4.2M
params
5.3min
train time

📊 訓練曲線

curves

📈 Test Per-attribute

attributeAPF1PRthrvalid
has_open 有門開啟0.99660.9770.9740.9790.761640
has_close 有門關閉0.97720.9200.9080.9320.321640

📥 模型下載 (R2 public)

🏗 模型架構

🧪 推論範例

img → model → [has_open=0.95, has_close=0.02]  → 圖中有開門
img → model → [has_open=0.02, has_close=0.98]  → 圖中有關門
img → model → [has_open=0.85, has_close=0.91]  → mixed(多扇門一開一關)
img → model → [has_open=0.10, has_close=0.15]  → 不確定(無門/看不清)
schema 設計:cvat 用單一 door tag + state select(open/closed/mixed/unknown),但訓練時轉成 2 個 binary head 處理 multi-door 場景。

📋 資料分布

splitopenclosedmixedunknowntotal
Train5181,8151,7424614,536
Validation101451539791,170
Test2873999541251,765

生成 2026-04-27| PPE 21-attr reportlabel guidesafety_rope proposal