簡介
物件偵測 = 找到影像中每個物件的位置（邊界框）+ 標籤（類別）。 YOLO（You Only Look Once）是王者👑－最快、最準確、最容易使用。
🎯 **為什麼選擇YOLO？ ** 即時（>100 FPS）、高精度（mAP 50%+）、1行程式碼即可偵測，Ultralytics生態系統極為成熟。
1.物體偵測101
1.1 分類、偵測、分割
Image Classification: "This is a cat" → 1 label
Object Detection: "Cat at (x,y,w,h)" → N bounding boxes + labels
Instance Segmentation: "Cat occupies these pixels" → Pixel-level masks
1.2 邊界框 — 圍繞物件的框
# Bounding box formats
# Format 1: (x_center, y_center, width, height) — YOLO format
bbox_yolo = [0.5, 0.5, 0.3, 0.4] # Normalized (0-1)
# Format 2: (x_min, y_min, x_max, y_max) — Pascal VOC format
bbox_voc = [150, 100, 350, 300] # Pixels
# Format 3: (x_min, y_min, width, height) — COCO format
bbox_coco = [150, 100, 200, 200] # Pixels
# Convert YOLO → VOC
def yolo_to_voc(bbox, img_w, img_h):
x_c, y_c, w, h = bbox
x_min = int((x_c - w/2) * img_w)
y_min = int((y_c - h/2) * img_h)
x_max = int((x_c + w/2) * img_w)
y_max = int((y_c + h/2) * img_h)
return [x_min, y_min, x_max, y_max]
1.3 IoU——並集的交集
"""IoU: đo mức overlap giữa 2 bounding boxes"""
def calculate_iou(box1, box2):
"""
box1, box2: [x_min, y_min, x_max, y_max]
"""
# Intersection
x_inter_min = max(box1[0], box2[0])
y_inter_min = max(box1[1], box2[1])
x_inter_max = min(box1[2], box2[2])
y_inter_max = min(box1[3], box2[3])
inter_area = max(0, x_inter_max - x_inter_min) * \
max(0, y_inter_max - y_inter_min)
# Union
area1 = (box1[2] - box1[0]) * (box1[3] - box1[1])
area2 = (box2[2] - box2[0]) * (box2[3] - box2[1])
union_area = area1 + area2 - inter_area
return inter_area / union_area if union_area > 0 else 0
# Ví dụ
pred_box = [100, 100, 300, 300]
gt_box = [120, 110, 310, 320]
iou = calculate_iou(pred_box, gt_box)
print(f"IoU: {iou:.4f}") # ~0.73
IoU Thresholds:
IoU > 0.5 → True Positive (mAP@50)
IoU > 0.75 → True Positive (mAP@75 — strict)
IoU < 0.5 → False Positive (miss!)
2. YOLO——你只看一次
2.1 核心思想
在YOLO之前，檢測使用2階段：提出區域→對每個區域進行分類（R-CNN，慢）。
YOLO：1-stage — 查看圖像僅一次 → 輸出所有框+類別。
Input Image (640×640)
↓
YOLO Backbone (feature extraction)
↓
YOLO Neck (feature fusion — FPN/PAN)
↓
YOLO Head (predict boxes + classes)
↓
NMS (Non-Max Suppression — lọc box trùng)
↓
Final Detections: [(class, confidence, x, y, w, h), ...]
2.2 YOLO的歷史
| 版本 | 年份 | 主要貢獻 | 地圖 (COCO) |
|---|---|---|---|
| YOLOv1 | 2016 | 2016最初的想法－1階段偵測 | 63.4 |
| YOLOv2 | 2017 | 2017錨框，批次標準化 | 78.6 |
| YOLOv3 | 2018 | 多尺度偵測，Darknet-53 | 33.0 (mAP@50:95) |
| YOLOv4 | 2020 | CSPDarknet、Mish 活化、Mosaic 增強 | 43.5 |
| YOLOv5 | 2020 | PyTorch，Ultralytics 生態系統，易於使用 | 48.2 |
| YOLOv6 | 2022 | 2022美團、BiC模組、SimOTA | 52.5 |
| YOLOv7 | 2022 | 2022 E-ELAN，模型重新參數化 | 56.8 |
| YOLOv8 | 2023 | 無錨、分離頭，Ultralytics | 53.9 |
| YOLOv9 | 2024 | 2024 PGI、GELAN架構 | 55.6 |
| YOLOv10 | 2024 | 2024免網管，效率驅動 | 54.4 |
| YOLOv11 | 2024 | C3k2 塊，注意，SOTA | 56.1 |
⭐ 實作： 使用 YOLOv8 或 YOLOv11 (Ultralytics)。最好的生態系統、清晰的文檔、最大的社群。
2.3 非極大值抑制（NMS）
"""NMS: loại bỏ bounding boxes trùng nhau"""
def nms(boxes, scores, iou_threshold=0.5):
"""
boxes: [[x1,y1,x2,y2], ...] — tất cả predicted boxes
scores: [0.9, 0.85, 0.7, ...] — confidence của mỗi box
"""
indices = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)
keep = []
while indices:
current = indices.pop(0)
keep.append(current)
remaining = []
for idx in indices:
iou = calculate_iou(boxes[current], boxes[idx])
if iou < iou_threshold: # Chỉ giữ box ít overlap
remaining.append(idx)
indices = remaining
return keep
3. 實踐：YOLO 與 Ultralytics
3.1 安裝
pip install ultralytics
3.2 推理－立即偵測
"""YOLO Detection — chỉ 3 dòng code!"""
from ultralytics import YOLO
# Load pretrained model (tự download)
model = YOLO("yolo11n.pt") # nano (nhanh nhất)
# model = YOLO("yolo11s.pt") # small
# model = YOLO("yolo11m.pt") # medium
# model = YOLO("yolo11l.pt") # large
# model = YOLO("yolo11x.pt") # extra large (chính xác nhất)
# Detect trên ảnh
results = model("street_photo.jpg")
# Hiển thị kết quả
results[0].show() # Mở ảnh với bounding boxes
results[0].save("output.jpg") # Lưu ảnh
3.3 分析詳細結果
"""Phân tích detection results"""
results = model("street.jpg")
for result in results:
boxes = result.boxes
for box in boxes:
# Bounding box coordinates
x1, y1, x2, y2 = box.xyxy[0].tolist()
# Confidence score
confidence = box.conf[0].item()
# Class
class_id = int(box.cls[0].item())
class_name = result.names[class_id]
print(f"📦 {class_name}: {confidence:.2%} "
f"at ({x1:.0f}, {y1:.0f}, {x2:.0f}, {y2:.0f})")
# Summary
print(f"\nTotal detections: {len(boxes)}")
print(f"Classes found: {set(result.names[int(c)] for c in boxes.cls)}")
3.4 影片偵測
"""Object detection trên video"""
import cv2
from ultralytics import YOLO
model = YOLO("yolo11n.pt")
# Mở video
cap = cv2.VideoCapture("traffic.mp4")
# Hoặc webcam: cap = cv2.VideoCapture(0)
# Video writer
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
fps = cap.get(cv2.CAP_PROP_FPS)
writer = cv2.VideoWriter("output.mp4", cv2.VideoWriter_fourcc(*"mp4v"), fps, (width, height))
while cap.isOpened():
ret, frame = cap.read()
if not ret:
break
# Detect
results = model(frame, verbose=False)
# Vẽ results lên frame
annotated = results[0].plot()
# Lưu
writer.write(annotated)
cap.release()
writer.release()
print("Done! Saved to output.mp4")
3.5 YOLO 任務：不只是偵測
"""YOLO hỗ trợ nhiều tasks"""
# Object Detection
det_model = YOLO("yolo11n.pt")
det_results = det_model("street.jpg")
# Instance Segmentation
seg_model = YOLO("yolo11n-seg.pt")
seg_results = seg_model("street.jpg")
# Pose Estimation
pose_model = YOLO("yolo11n-pose.pt")
pose_results = pose_model("person.jpg")
# Classification
cls_model = YOLO("yolo11n-cls.pt")
cls_results = cls_model("cat.jpg")
# OBB (Oriented Bounding Boxes)
obb_model = YOLO("yolo11n-obb.pt")
obb_results = obb_model("aerial.jpg")
4. 評估指標
4.1 精確度、召回率、mAP
Predicted Positive Predicted Negative
Actual Positive TP (True Positive) FN (False Negative)
Actual Negative FP (False Positive) TN (True Negative)
Precision = TP / (TP + FP) → "Trong tất cả detect, bao nhiêu đúng?"
Recall = TP / (TP + FN) → "Trong tất cả objects thật, bao nhiêu detect được?"
4.2 mAP（平均精確度）
"""Đánh giá YOLO model"""
from ultralytics import YOLO
model = YOLO("yolo11n.pt")
# Evaluate trên COCO val set
metrics = model.val(data="coco.yaml")
print(f"mAP@50: {metrics.box.map50:.4f}") # mAP at IoU=0.5
print(f"mAP@50:95: {metrics.box.map:.4f}") # mAP at IoU=0.5:0.95
print(f"Precision: {metrics.box.mp:.4f}") # Mean Precision
print(f"Recall: {metrics.box.mr:.4f}") # Mean Recall
4.3 YOLO模型比較
| 型號 | 參數 | 地圖@50 | 地圖@50:95 | 速度（T4 GPU） |
|---|---|---|---|---|
| YOLOv11n | 260 萬 | 70.3 | 70.3 39.5 | 39.5 1.5 毫秒 |
| YOLOv11s | 940 萬 | 77.1 | 77.1 47.0 | 47.0 2.5 毫秒 |
| YOLOv11m | 2010 萬 | 80.4 | 80.4 51.5 | 51.5 4.7 毫秒 |
| YOLOv11l | 2530 萬 | 81.2 | 81.2 53.4 | 53.4 6.2 毫秒 |
| YOLOv11x | 56.9M | 82.0 | 54.7 | 54.7 11.3 毫秒 |
5. 進階YOLO配置
"""Cấu hình detection chi tiết"""
results = model.predict(
source="image.jpg", # ảnh, video, folder, url, webcam
conf=0.25, # Confidence threshold (default 0.25)
iou=0.45, # NMS IoU threshold (default 0.7)
classes=[0, 1, 2], # Chỉ detect classes cụ thể (0=person, 1=bicycle, 2=car)
max_det=300, # Số detection tối đa
imgsz=640, # Input size
device="cuda", # GPU
save=True, # Lưu ảnh kết quả
save_txt=True, # Lưu labels txt
save_conf=True, # Lưu confidence trong txt
show=False, # Hiển thị real-time
verbose=False, # Tắt log
)
總結
| 概念 | 記住 |
|---|---|
| 物体检测 | 查找对象的位置（bbox）+标签（类） |
| YOLO | 1 級偵測器，即時，SOTA |
| 欠条 | 测量预测框和真实框之间的重叠 |
| 网络管理系统 | 消除重复盒子，保留最好的盒子 |
| 地图 | 检测的主要指标：平均精度 |
| 超解 | YOLO 的 Python 库 — 3 行检测代码 |
一般練習
- 快速入门： 安装 Ultralytics，检测 10 个不同的图像（街道、室内、自然）。統計每個班級。
- 視訊偵測： 偵測視訊流量。平均有多少輛車/車架？
- 模型大小比较： 比较 yolo11n、yolo11m 和 yolo11x：准确性、速度、内存。
- IoU计算器： 实现IoU函数，用5对不同的盒子进行测试。
- 自定义过滤器： 编写一个脚本，仅以 > 80% 的置信度检测 人（0 类）。
下一篇文章： YOLO 自定义训练 - 标记您自己的数据，针对特定问题训练您自己的模型。