forklift v20260525 report

⚠️ Export Bug — v525 訓練資料不完整（NOT for production deploy）

症狀：14 個自 v519 後新 acceptance 的 cvat #9 任務（IRODA_WAREHOUSE 系列 CH04/CH05_2026051x_001 等）frames 幾乎全沒進 training：

task	cvat size	實際進 val/test
4867 IRODA_WAREHOUSE_CH04_20260514_001	235 frames	val 1 img
4902 IRODA_WAREHOUSE_CH04_20260506_001	91 frames	val 1 img
4869 IRODA_WAREHOUSE_CH05_20260513_001	156 frames	0 ❌
4900 IRODA_WAREHOUSE_CH05_20260506_001	175 frames	0 ❌
4862 IRODA_WAREHOUSE_CH04_20260515_001	1060 frames (Train)	0 ❌

根因：v519/v525 export script 把 cvat meta["frames"] 當 image list iterate，但新 IRODA 任務是 video mode（單一 mp4 entry，例 頻道4-倉庫進料區_20260514_928.mp4）。export 對每 task 只看到 1 個 entry，且 disk resolver 找 mp4 找不到 frame image → 直接 skip 整 task。

影響：v525 ≈ v519 同 dataset 重訓；本報告上方 +3.7pp Precision 是 retrain noise，不是新資料貢獻。Per-source 區段（下方）的 IRODA bucket 完全沒出現可佐證。

跟標記師討論點：cvat #9 是要用 video mode 上傳？還是希望 export 端負責把 mp4 拆 frame？PPE export (export_p12_v521.py) 已有 video frame 處理可參考。

下一步：(1) 修 export 加 video frame 處理 (2) 重 export + 開 v526 訓練 (3) v525 best.pt 暫不切 production，繼續用 v519

🚜 Forklift YOLO Detection v20260525

train @ 5090-2 GPU 1, single-card batch=64, 2026-05-25 cvat #9 export (87 accepted tasks, 11,431 train imgs)

📊 Key Metrics — v20260525 vs v20260519

Metric	v20260519 (ep100)	v20260525 best (ep98)	Δ
Precision	0.949	0.98595	+3.7pp
Recall	0.907	0.90444	-0.3pp
mAP50	0.953	0.94962	-0.3pp
mAP50-95	0.820	0.81669	-0.3pp

Metric

v20260519 (ep100)

v20260525 best (ep98)

Precision

0.949

0.98595

+3.7pp

Recall

0.907

0.90444

-0.3pp

mAP50

0.953

0.94962

-0.3pp

mAP50-95

0.820

0.81669

-0.3pp

🛡️ SOP v20260518 補強

1. 全 project shape distribution (accepted tasks only)

shape type	count
rectangle	10625
polygon	8154
polyline	1

shape type

count

rectangle

10625

polygon

8154

polyline

無 rectangle（v518 翻車主因），全 polygon → 走 polygon→bbox (min/max xy)。symmetry with v519 fix.

conf	P	R	mAP50	mAP50-95
0.001	0.9869	0.9042	0.9492	0.8165
0.01	0.9869	0.9042	0.9561	0.8331
0.05	0.9869	0.9042	0.9594	0.8407
0.1	0.9869	0.9042	0.9600	0.8425
0.15	0.9869	0.9042	0.9593	0.8427
0.2	0.9869	0.9042	0.9573	0.8427
0.25	0.9869	0.9042	0.9572	0.8431
0.3	0.9869	0.9042	0.9568	0.8432
0.35	0.9869	0.9042	0.9559	0.8434
0.4	0.9869	0.9042	0.9526	0.8421
0.45	0.9869	0.9042	0.9521	0.8421
0.5	0.9869	0.9042	0.9502	0.8417
0.55	0.9869	0.9046	0.9493	0.8419
0.6	0.9887	0.9012	0.9477	0.8409
0.65	0.9905	0.8926	0.9435	0.8384
0.7	0.9913	0.8840	0.9393	0.8356
0.75	0.9922	0.8754	0.9351	0.8331
0.8	0.9931	0.8643	0.9296	0.8304
0.85	0.9938	0.8282	0.9118	0.8205
0.9	0.9962	0.6813	0.8390	0.7758
0.95	0.9979	0.4158	0.7068	0.6761

conf

mAP50

mAP50-95

0.001

0.9869

0.9042

0.9492

0.8165

0.01

0.9869

0.9042

0.9561

0.8331

0.05

0.9869

0.9042

0.9594

0.8407

0.1

0.9869

0.9042

0.9600

0.8425

0.15

0.9869

0.9042

0.9593

0.8427

0.2

0.9869

0.9042

0.9573

0.8427

0.25

0.9869

0.9042

0.9572

0.8431

0.3

0.9869

0.9042

0.9568

0.8432

0.35

0.9869

0.9042

0.9559

0.8434

0.4

0.9869

0.9042

0.9526

0.8421

0.45

0.9869

0.9042

0.9521

0.8421

0.5

0.9869

0.9042

0.9502

0.8417

0.55

0.9869

0.9046

0.9493

0.8419

0.6

0.9887

0.9012

0.9477

0.8409

0.65

0.9905

0.8926

0.9435

0.8384

0.7

0.9913

0.8840

0.9393

0.8356

0.75

0.9922

0.8754

0.9351

0.8331

0.8

0.9931

0.8643

0.9296

0.8304

0.85

0.9938

0.8282

0.9118

0.8205

0.9

0.9962

0.6813

0.8390

0.7758

0.95

0.9979

0.4158

0.7068

0.6761

source	n_imgs	P	R	mAP50	mAP50-95
other	1148	0.9867	0.9040	0.9498	0.8181
HONCHUAN	23	1.0000	0.9241	0.9472	0.6898

source

n_imgs

mAP50

mAP50-95

other

1148

0.9867

0.9040

0.9498

0.8181

HONCHUAN

1.0000

0.9241

0.9472

0.6898

source	n_imgs	P	R	mAP50	mAP50-95
other	2157	0.9897	0.9817	0.9940	0.8721
HONCHUAN	107	0.9661	1.0000	0.9937	0.9123

source

n_imgs

mAP50

mAP50-95

other

2157

0.9897

0.9817

0.9940

0.8721

HONCHUAN

107

0.9661

1.0000

0.9937

0.9123

{ "epochs": 100, "patience": 30, "batch": 64, "imgsz": 640, "device": "1 (single GPU)", "model_base": "yolo26n.pt", "optimizer": "auto -> MuSGD lr=0.01 mom=0.9", "augmentation": "default (mosaic=1.0, close_mosaic=10, fliplr=0.5, hsv_v=0.4)", "fully_inherited_from": "v20260519" }

📦 Dataset

cvat #9 forklift_detection — 87 acceptance+completed tasks (+14 vs v519)

Train: 11,431 imgs (1,202 backgrounds) | Val: 1,171 (139 bg) | Test: see test split

14 new acc tasks since v519: 3 empty (Train negative, operator confirmed normal)

NEW task sources: IRODA_WAREHOUSE_CH04/05_2026051x 系列, HONCHUAN_UNKNOWN_CH00_20260504_041

⚠️ Export Bug — v525 訓練資料不完整（NOT for production deploy）

🚜 Forklift YOLO Detection v20260525

📊 Key Metrics — v20260525 vs v20260519

🛡️ SOP v20260518 補強

1. 全 project shape distribution (accepted tasks only)

2. Recall @ Precision=0.95 (SOP primary metric)

3. Per-source metrics（domain shift audit）

val

test

🔄 Training Curve

📈 PR / F1 Curves

🔍 Confusion Matrix

👀 Val predictions sample

⚙️ Hyperparams (full inheritance from v20260519)

📦 Dataset