模型训练 (使用训练好的模型)

• Ultralytics YOLO26 完整训练指南

🏞 单张图片推理

from ultralytics import YOLO

model = YOLO("runs/train/my_yolo26_exp/weights/best.pt")

results = model("path/to/test_image.jpg")

for result in results:
    boxes = result.boxes
    for box in boxes:
        cls_id = int(box.cls[0])             # 类别 ID
        cls_name = result.names[cls_id]       # 类别名称
        confidence = float(box.conf[0])       # 置信度
        x1, y1, x2, y2 = box.xyxy[0].tolist()  # 边界框坐标
        print(f"检测到: {cls_name}, 置信度: {confidence:.2f}, 坐标: ({x1:.0f},{y1:.0f})-({x2:.0f},{y2:.0f})")

    result.show()   # 显示结果图片
    result.save()   # 保存结果到 runs/detect/predict/

🗂 批量图片推理

from ultralytics import YOLO

model = YOLO("runs/train/my_yolo26_exp/weights/best.pt")

results = model(
    source="path/to/test_images/",   # 图片目录
    conf=0.5,                         # 置信度阈值
    iou=0.45,                         # IoU 阈值（使用 one-to-many 头部时）
    save=True,                        # 保存结果图片
    save_txt=True,                    # 保存检测结果 txt
)

🎞 视频推理

from ultralytics import YOLO

model = YOLO("runs/train/my_yolo26_exp/weights/best.pt")

results = model(
    source="path/to/video.mp4",
    conf=0.5,
    save=True,       # 保存标注后的视频
    stream=True,     # 流式处理，节省内存
)

for result in results:
    pass  # 逐帧处理

📹 摄像头实时推理

from ultralytics import YOLO

model = YOLO("runs/train/my_yolo26_exp/weights/best.pt")

results = model(
    source=0,        # 0 表示默认摄像头
    conf=0.5,
    show=True,       # 实时显示窗口
    stream=True,
)

for result in results:
    pass

⌨️ 命令行推理

# 单张图片
yolo detect predict model=runs/train/my_yolo26_exp/weights/best.pt source=test.jpg conf=0.5

# 图片目录
yolo detect predict model=runs/train/my_yolo26_exp/weights/best.pt source=test_images/ conf=0.5 save=True

# 视频
yolo detect predict model=runs/train/my_yolo26_exp/weights/best.pt source=video.mp4 conf=0.5 save=True

↔️ 双头架构切换

YOLO26 支持两种推理头：

model = YOLO("runs/train/my_yolo26_exp/weights/best.pt")

# One-to-One 头（默认）：端到端，无需 NMS，速度更快
results = model.predict("image.jpg")

# One-to-Many 头：需要 NMS 后处理，精度略高
results = model.predict("image.jpg", end2end=False)