\ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 YOLO(You Only Look Once) \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \ub2e8\uc77c \ubaa8\ub378\ub85c \uac1d\uccb4 \uac80\ucd9c\uacfc \ubd84\ub958\ub97c \ub3d9\uc2dc\uc5d0 \uc218\ud589\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc124\uba85\ud569\ub2c8\ub2e4. YOLO\ub294 \uc2e4\uc2dc\uac04 \uac1d\uccb4 \uac80\ucd9c\uc5d0 \ucd5c\uc801\ud654\ub41c \ub525\ub7ec\ub2dd \ubaa8\ub378\ub85c, \ud55c \ubc88\uc758 \ud328\uc2a4\uc5d0\uc11c \ubcf5\uc7a1\ud55c \uc815\ubcf4 \ucc98\ub9ac\ub97c \uac00\ub2a5\ud558\uac8c \ud558\uc5ec \ub9e4\uc6b0 \ud6a8\uc728\uc801\uc785\ub2c8\ub2e4.<\/p>\n
YOLO\ub294 \ub2e4\uc591\ud55c \uac1d\uccb4\ub97c \ud558\ub098\uc758 \uc774\ubbf8\uc9c0\uc5d0\uc11c \ub3d9\uc2dc\uc5d0 \uac10\uc9c0\ud558\uace0 \ubd84\ub958\ud558\ub294 \ub2a5\ub825\uc744 \uac00\uc9c0\uace0 \uc788\uc73c\uba70, \uc774 \ub54c\ubb38\uc5d0 \uba40\ud2f0\ud0dc\uc2a4\ud0b9 \ubaa8\ub378\uc744 \uad6c\ucd95\ud558\ub294 \ub370 \uc801\ud569\ud55c \uc120\ud0dd\uc785\ub2c8\ub2e4. YOLO\uc758 \uc791\ub3d9 \ubc29\uc2dd\uc740 \uc774\ubbf8\uc9c0 \uc804\uccb4\ub97c \uadf8\ub9ac\ub4dc\ub85c \ub098\ub204\uace0 \uac01 \uadf8\ub9ac\ub4dc\uc5d0\uc11c \uac1d\uccb4\uc758 \uc874\uc7ac \uc5ec\ubd80\uc640 \uadf8 \uac1d\uccb4\uc758 \ud074\ub798\uc2a4 \ud655\ub960\uc744 \uc608\uce21\ud558\ub294 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
YOLO\uc758 \uae30\ubcf8 \uad6c\uc870\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uc694\uc18c\ub85c \uc774\ub8e8\uc5b4\uc838 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
YOLO \ubaa8\ub378\uc744 \uad6c\ucd95\ud558\uace0 \ud6c8\ub828\ud558\uae30 \uc704\ud574\uc11c\ub294 \uba87 \uac00\uc9c0 \ud544\uc218 \ud328\ud0a4\uc9c0\ub97c \uc124\uce58\ud574\uc57c \ud569\ub2c8\ub2e4. \ub2e4\uc74c\uc740 Python \ud658\uacbd\uc5d0\uc11c YOLOv5\ub97c \uc0ac\uc6a9\ud558\ub294 \uc608\uc2dc\uc785\ub2c8\ub2e4.<\/p>\n
!pip install torch torchvision torchaudio\n!pip install matplotlib\n!pip install opencv-python\n!pip install numpy<\/code><\/pre>\n\ub370\uc774\ud130 \uc900\ube44<\/h2>\n
\uba40\ud2f0\ud0dc\uc2a4\ud0b9 \ubaa8\ub378\uc744 \uad6c\ucd95\ud558\uae30 \uc704\ud574\uc11c\ub294 \uac1d\uccb4 \uac80\ucd9c\uacfc \ubd84\ub958\uc5d0 \ud544\uc694\ud55c \ub370\uc774\ud130\ub97c \uc900\ube44\ud574\uc57c \ud569\ub2c8\ub2e4. COCO \ub370\uc774\ud130\uc14b\uc774\ub098 Pascal VOC \ub370\uc774\ud130\uc14b\uacfc \uac19\uc740 \uacf5\uac1c \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub370\uc774\ud130\ub294 \ub450 \uac00\uc9c0 \ud615\uc2dd\uc73c\ub85c \uc900\ube44\ud574\uc57c \ud569\ub2c8\ub2e4:<\/p>\n
\n- \ubb3c\uccb4\uc758 \uc704\uce58\ub97c \ub098\ud0c0\ub0b4\ub294 \ubc14\uc6b4\ub529 \ubc15\uc2a4 \uc88c\ud45c<\/li>\n
- \uac01 \ubb3c\uccb4\uc5d0 \ud574\ub2f9\ud558\ub294 \ud074\ub798\uc2a4 \ub808\uc774\ube14<\/li>\n<\/ul>\n
\ub370\uc774\ud130\uc14b \ud3ec\ub9f7<\/h3>\n
YOLO\uc5d0\uc11c \uc0ac\uc6a9\ud558\ub294 \ub370\uc774\ud130\uc14b \ud3ec\ub9f7\uc740 \uac01 \uac1d\uccb4\uc5d0 \ub300\ud55c \uc815\ubcf4\ub97c \ud3ec\ud568\ud558\ub294 \ud14d\uc2a4\ud2b8 \ud30c\uc77c\uc758 \ud615\ud0dc\uc785\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \ub2e4\uc74c\uacfc \uac19\uc740 \uc815\ubcf4\ub97c \ud3ec\ud568\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
0 0.5 0.5 0.2 0.2 # \ud074\ub798\uc2a4 \ubc88\ud638, x_center, y_center, width, height\n1 0.4 0.4 0.1 0.1 # \ub610 \ub2e4\ub978 \uac1d\uccb4 \uc815\ubcf4<\/code><\/pre>\nYOLO \ubaa8\ub378 \uc544\ud0a4\ud14d\ucc98<\/h2>\n
YOLO \ubaa8\ub378\uc758 \uc544\ud0a4\ud14d\ucc98 \uad6c\uc131 \uc694\uc18c\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- Backbone:<\/strong> \ud2b9\uc9d5 \ucd94\ucd9c<\/li>\n
- Neck:<\/strong> \ub2e4\uc591\ud55c \uacc4\uce35\uc5d0\uc11c\uc758 \ud2b9\uc9d5\uc744 \uacb0\ud569\ud558\uc5ec \ub354 \uac15\ub825\ud55c \ud2b9\uc9d5\uc744 \uc0dd\uc131<\/li>\n
- Head:<\/strong> \ucd5c\uc885\uc801\uc73c\ub85c \uac10\uc9c0\ud55c \uac1d\uccb4\uc758 \ubc14\uc6b4\ub529 \ubc15\uc2a4 \ubc0f \ud074\ub798\uc2a4 \uc608\uce21<\/li>\n<\/ul>\n
YOLOv5 \ubaa8\ub378 \uad6c\ud604<\/h3>\n
YOLOv5\ub294 PyTorch\ub85c \uad6c\ud604\ub41c \ucd5c\uc2e0 YOLO \ubc84\uc804 \uc911 \ud558\ub098\uc785\ub2c8\ub2e4. \ub2e4\uc74c\uc740 PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec YOLOv5 \ubaa8\ub378\uc744 \ubd88\ub7ec\uc624\ub294 \ucf54\ub4dc \uc608\uc2dc\uc785\ub2c8\ub2e4:<\/p>\n
import torch\n\n# YOLOv5 \ubaa8\ub378 \ub85c\ub4dc\nmodel = torch.hub.load('ultralytics\/yolov5', 'yolov5s', pretrained=True)\n<\/code><\/pre>\n\ubaa8\ub378 \ud6c8\ub828<\/h2>\n
Yolov5 \ubaa8\ub378\uc744 \ucee4\uc2a4\ud130\ub9c8\uc774\uc988\ud558\uc5ec \ud6c8\ub828\ud558\uae30 \uc704\ud574\uc11c\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \ubc29\ubc95\uc73c\ub85c \uc9c4\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
!python train.py --img 640 --batch 16 --epochs 50 --data data.yaml --weights yolov5s.pt<\/code><\/pre>\n\uc704 \ucf54\ub4dc\ub97c \uc0ac\uc6a9\ud558\uba74 640×640 \ud574\uc0c1\ub3c4\uc758 \uc774\ubbf8\uc9c0\ub97c \uc0ac\uc6a9\ud558\uc5ec, \ubc30\uce58 \ud06c\uae30 16\uc73c\ub85c 50 epoch \ud6c8\ub828\ud558\uac8c \ub429\ub2c8\ub2e4. ‘data.yaml’ \ud30c\uc77c\uc740 \ud559\uc2b5\uc2dc \uc0ac\uc6a9\ud560 \ub370\uc774\ud130\uc14b\uc5d0 \ub300\ud55c \uc815\ubcf4\ub97c \ud3ec\ud568\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
\ud14c\uc2a4\ud2b8\uc640 \ud3c9\uac00<\/h2>\n
\ubaa8\ub378 \ud6c8\ub828 \ud6c4, YOLOv5 \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc774\ubbf8\uc9c0\ub098 \ube44\ub514\uc624 \uc2a4\ud2b8\ub9bc\uc5d0\uc11c \uac1d\uccb4\ub97c \ud0d0\uc9c0\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c\uc740 \uc774\ubbf8\uc9c0\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud14c\uc2a4\ud2b8\ud558\ub294 \uc608\uc2dc\uc785\ub2c8\ub2e4:<\/p>\n
results = model('test.jpg') # \uc785\ub825 \uc774\ubbf8\uc9c0\nresults.show() # \uacb0\uacfc \uc774\ubbf8\uc9c0 \ud45c\uc2dc<\/code><\/pre>\n\uc815\ud655\ub3c4 \ud3c9\uac00<\/h3>\n
\ud6c8\ub828\ub41c \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\uae30 \uc704\ud574 Mean Average Precision (mAP) \ub4f1\uc758 \uc9c0\ud45c\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc778 \ud3c9\uac00 \ubc29\ubc95\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
from utils.metrics import box_iou\n\n# \uc608\uce21 \ubc15\uc2a4 \ubc0f \uc2e4\uc81c \ubc15\uc2a4 \uc608\uc81c\npred_boxes = torch.tensor([[0.5, 0.5, 0.2, 0.2]])\ntrue_boxes = torch.tensor([[0.5, 0.5, 0.1, 0.1]])\n\niou = box_iou(pred_boxes, true_boxes)\nprint(f'IOU: {iou}') # IOU \uac12 \ucd9c\ub825<\/code><\/pre>\n\uba40\ud2f0\ud0dc\uc2a4\ud0b9\uc744 \ud1b5\ud55c \ub300\uc548<\/h2>\n
YOLO\ub97c \uc0ac\uc6a9\ud558\uc5ec \ub3d9\uc2dc\uc5d0 \uc5ec\ub7ec \uc791\uc5c5(\uac1d\uccb4 \uac80\ucd9c\uacfc \ubd84\ub958)\uc744 \uc2e4\ud589\ud560 \uc218 \uc788\ub294 \uba40\ud2f0\ud0dc\uc2a4\ud0b9 \ubaa8\ub378\uc744 \ub9cc\ub4e4\uc5b4 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740 \uc785\ub825\ub41c \uc774\ubbf8\uc9c0\uc5d0\uc11c \uac1d\uccb4\ub97c \uac10\uc9c0\ud558\uace0, \uac01 \uac1d\uccb4\uc758 \ubd84\ub958\ub97c \ub3d9\uc2dc\uc5d0 \uc218\ud589\ud569\ub2c8\ub2e4. \ub530\ub77c\uc11c \uac01 \uac1d\uccb4\uc5d0 \ub300\ud574 \ube60\ub974\uace0 \ud6a8\uc728\uc801\uc778 \uac80\ucd9c\uacfc \ubd84\ub958\uac00 \uac00\ub2a5\ud569\ub2c8\ub2e4.<\/p>\n
\uba40\ud2f0\ud0dc\uc2a4\ud0b9 \ubaa8\ub378 \uc810\uac80<\/h3>\n
YOLO\uc758 \uae30\ubcf8 \uad6c\uc870\ub97c \uae30\ubc18\uc73c\ub85c \uba40\ud2f0\ud0dc\uc2a4\ud0b9 \ub2a5\ub825\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud55c \ubc29\ubc95\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- \ucd9c\ub825\uce35\uc744 \uc218\uc815\ud558\uc5ec \ud074\ub798\uc2a4\uc640 \ubc14\uc6b4\ub529\ubc15\uc2a4\ub97c \ub3d9\uc2dc\uc5d0 \uc608\uce21\ud558\ub3c4\ub85d \ud55c\ub2e4.<\/li>\n
- \ud6c8\ub828 \ub370\uc774\ud130\ub97c \ud1b5\ud55c \uac1d\uccb4 \ubd84\ub958 \uc815\ubcf4\ub97c \ub124\ud2b8\uc6cc\ud06c\uc5d0 \ud559\uc2b5\uc2dc\ud0a8\ub2e4.<\/li>\n<\/ul>\n
\ubaa8\ub378 \ubcc0\ud615<\/h3>\n
\uba40\ud2f0\ud0dc\uc2a4\ud0b9 \ub2a5\ub825\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud574 CustomHead \ubaa8\ub4c8\uc744 \uc124\uacc4\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \ubaa8\ub4c8\uc740 \uac1d\uccb4\uc758 \ubc14\uc6b4\ub529 \ubc15\uc2a4\uc640 \ubd84\ub958 \uc815\ubcf4 \ubaa8\ub450\ub97c \ucd9c\ub825\ud558\ub3c4\ub85d \uad6c\uc131\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
import torch.nn as nn\n\nclass CustomHead(nn.Module):\n def __init__(self):\n super(CustomHead, self).__init__()\n self.fc1 = nn.Linear(1024, 512) # Feature Extraction\n self.box_output = nn.Linear(512, 4) # Bounding Box Output\n self.class_output = nn.Linear(512, num_classes) # Class Output\n\n def forward(self, x):\n x = nn.ReLU()(self.fc1(x))\n boxes = self.box_output(x)\n classes = self.class_output(x)\n return boxes, classes<\/code><\/pre>\n\uacb0\ub860<\/h2>\n
\ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 YOLO \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \uba40\ud2f0\ud0dc\uc2a4\ud0b9 \ubaa8\ub378\uc744 \uad6c\ucd95\ud558\ub294 \ubc29\ubc95\uc744 \uc124\uba85\ud588\uc2b5\ub2c8\ub2e4. YOLO\ub294 \uac1d\uccb4 \uac80\ucd9c \ubc0f \ubd84\ub958\ub97c \ub3d9\uc2dc\uc5d0 \uc218\ud589\ud560 \uc218 \uc788\ub294 \ub2a5\ub825\uc744 \uc81c\uacf5\ud558\uba70, \uc774\ub97c \ud1b5\ud574 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud65c\uc6a9\ud560 \uc218 \uc788\ub294 \ud6a8\uc728\uc801\uc778 \ubaa8\ub378\uc744 \ub9cc\ub4e4 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uba40\ud2f0\ud0dc\uc2a4\ud0b9 \ubaa8\ub378\uc744 \uad6c\ucd95\ud558\ub294 \uacfc\uc815\uc5d0\uc11c YOLO\uc758 \ub0b4\ubd80 \uad6c\uc870\uc640 \ud2b9\uc9d5, \ud6c8\ub828 \ubc29\ubc95 \ubc0f \ud3c9\uac00 \ubc29\ubc95\uc5d0 \ub300\ud574 \uae4a\uc774 \uc774\ud574\ud560 \uc218 \uc788\uc5c8\uc2b5\ub2c8\ub2e4.<\/p>\n
\ucc38\uace0\uc790\ub8cc<\/h2>\n