🔒 Safety Rope Detection — v20260502

RoI Align MobileNetV3-L 2-class 訓練日期：2026-04-28 ｜ 來源：cvat2 project 8 (safety_rope_detection)

📊 主要指標

📐 模型架構

📈 訓練曲線

🎯 Test Confusion Matrix（thr=0.075）

🚀 模型具體用法

① PyTorch 推論程式

import torch, torch.nn as nn
import torchvision.ops as tvops
import timm
import numpy as np
from PIL import Image
from ultralytics import YOLO

# ── Model ──────────────────────────────────────────────
class RoISafetyRopeModel(nn.Module):
    def __init__(self, backbone_name="mobilenetv3_large_100.ra_in1k"):
        super().__init__()
        self.backbone = timm.create_model(backbone_name, pretrained=False, features_only=True)
        ch = self.backbone.feature_info.channels()[-1]   # 960
        self.roi_align = tvops.RoIAlign((7,7), spatial_scale=1/32, sampling_ratio=2)
        self.head = nn.Sequential(
            nn.Conv2d(ch, 256, 3, padding=1), nn.ReLU(inplace=True),
            nn.AdaptiveAvgPool2d((1,1)), nn.Flatten(),
            nn.Dropout(0.3), nn.Linear(256, 2),
        )
    def forward(self, image, rois):
        feats = self.backbone(image)[-1]
        return self.head(self.roi_align(feats, rois))

# ── 載入 ──────────────────────────────────────────────
device = torch.device("cuda")
ckpt = torch.load("safety_rope_v20260502_best.pt", map_location=device, weights_only=False)
model = RoISafetyRopeModel(ckpt["backbone_name"]).to(device).eval()
model.load_state_dict(ckpt["model_state"])
yolo = YOLO("person_seg_yolo11n.pt")  # person detector
THR = ckpt["thr"]              # 0.0749
IMG_SIZE = ckpt["img_size"]    # 640
EX = ckpt["expand_x"]          # 1.0
EYT = ckpt["expand_y_top"]     # 0.2
EYB = ckpt["expand_y_bot"]     # 1.5

# ── 預測（單張圖） ────────────────────────────────────
mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
std  = np.array([0.229, 0.224, 0.225], dtype=np.float32)

def predict(img_pil):
    W, H = img_pil.size
    persons = yolo(img_pil, verbose=False, conf=0.35)[0].boxes.xyxy.cpu().numpy()
    if len(persons) == 0:
        return []
    img_resized = img_pil.resize((IMG_SIZE, IMG_SIZE))
    arr = (np.array(img_resized, dtype=np.float32)/255.0 - mean) / std
    x = torch.from_numpy(arr.transpose(2,0,1)).unsqueeze(0).to(device)
    sx, sy = IMG_SIZE/W, IMG_SIZE/H
    rois = []
    for x1, y1, x2, y2 in persons:
        w, h = x2-x1, y2-y1
        ex1 = max(0, x1 - w*EX); ey1 = max(0, y1 - h*EYT)
        ex2 = min(W, x2 + w*EX); ey2 = min(H, y2 + h*EYB)
        rois.append([0.0, ex1*sx, ey1*sy, ex2*sx, ey2*sy])
    rois_t = torch.tensor(rois, dtype=torch.float32).to(device)
    with torch.no_grad():
        probs = torch.softmax(model(x, rois_t), dim=-1)[:,1].cpu().numpy()
    return [
        {"bbox": persons[i].tolist(),
         "prob_correct": float(probs[i]),
         "predicted": "correct" if probs[i] >= THR else "wrong"}
        for i in range(len(persons))
    ]

② Live RTSP / Webcam 整合

已整合到 rai-model-viewer，部署於 http://192.168.53.21:7860/。 選 model = safety_rope_v20260502 (安全背繩) 後，每個 person 上會疊 bbox 與 CORRECT/WRONG 標籤， 右側直方圖即時顯示每人的安全繩信心值。

③ Threshold 調整建議

📦 資料集

📥 模型下載

R2 Bucket：rai-models / safety_rope_v20260502 / best.pt
Public URL（待上傳）：https://pub-478929a98a5c440cb22c2241c0bde314.r2.dev/safety_rope_v20260502/best.pt

🛠️ 訓練指令

python3 train_safety_rope_roi.py v20260502 \
  --manifest /home/ubuntu/factory_ppe/safety_rope_dataset/manifest_v1.csv \
  --backbone mobilenetv3_large_100.ra_in1k \
  --batch 32 --epochs 30 --lr 3e-4 --wd 0.01 --patience 6

📋 Hyperparameters

Generated 2026-04-28 ｜ 回到目錄