YOLO

YOLO-র মূল ধারণা

• ইমেজকে S×S গ্রিডে ভাগ — প্রতিটি cell B টি box predict করে।
• প্রতিটি box: (x, y, w, h, confidence) + class probabilities।
• Single network → end-to-end → real-time।
• NMS দিয়ে overlapping box ছাঁকা।

সংস্করণ-এ ক্রমবিকাশ

Ultralytics দিয়ে দ্রুত শুরু

pip install ultralytics

from ultralytics import YOLO
model = YOLO("yolov8n.pt")            # nano — দ্রুত
results = model("street.jpg")
results[0].show()
results[0].save("out.jpg")

নিজের ডেটায় Train (Custom)

# data.yaml
path: ./dataset
train: images/train
val: images/val
names: ["helmet", "no_helmet"]

# label ফাইল (per image .txt):
#   class_id x_center y_center width height   (সব normalized)

model = YOLO("yolov8s.pt")
model.train(data="data.yaml", epochs=50, imgsz=640, batch=16)
model.val()
model.predict("test.jpg", save=True)

Multi-task — Detect, Segment, Pose, Classify

YOLO("yolov8n.pt")        # detection
YOLO("yolov8n-seg.pt")    # instance segmentation
YOLO("yolov8n-pose.pt")   # keypoint / pose
YOLO("yolov8n-cls.pt")    # classification

Real-time Webcam

import cv2
from ultralytics import YOLO

model = YOLO("yolov8n.pt")
cap = cv2.VideoCapture(0)
while cap.isOpened():
    ok, frame = cap.read()
    if not ok: break
    res = model(frame, verbose=False)
    cv2.imshow("YOLO", res[0].plot())
    if cv2.waitKey(1) == ord("q"): break
cap.release()

Export — Deployment

model.export(format="onnx")        # ONNX
model.export(format="engine")      # TensorRT
model.export(format="coreml")      # iOS
model.export(format="tflite")      # Mobile

অনুশীলন

সারসংক্ষেপ