YOLO(You Only Look Once) \ubaa8\ub378\uc740 \uac1d\uccb4 \ud0d0\uc9c0 \ubd84\uc57c\uc5d0\uc11c \uac00\uc7a5 \uc778\uae30 \uc788\ub294 \uae30\uc220 \uc911 \ud558\ub098\uc785\ub2c8\ub2e4. YOLO\ub294 \uc785\ub825 \uc774\ubbf8\uc9c0 \uc804\uccb4\ub97c \ud55c \ubc88\uc758 \uc804\ubc29\ud5a5 \ud328\uc2a4\uc5d0\uc11c \ucc98\ub9ac\ud558\uc5ec \uc774\ubbf8\uc9c0\ub97c \ubd84\uc11d\ud558\ubbc0\ub85c, \uc18d\ub3c4\uac00 \ube60\ub974\uace0 \ud6a8\uc728\uc801\uc785\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \uc2e4\uc81c \ud504\ub85c\uc81d\ud2b8\uc5d0\uc11c YOLO\uc758 \uc131\ub2a5\uc744 \ucd5c\uc801\ud654\ud558\uae30 \uc704\ud574\uc11c\ub294 \ub2e4\uc591\ud55c \ud30c\ub77c\ubbf8\ud130 \ud29c\ub2dd\uc774 \ud544\uc694\ud569\ub2c8\ub2e4.<\/p>\n
YOLO \ubaa8\ub378\uc740 \uc5ec\ub7ec \uac00\uc9c0 \uc8fc\uc694 \ud30c\ub77c\ubbf8\ud130\ub97c \uac00\uc9c0\uace0 \uc788\uc73c\uba70, \uc774\ub4e4\uc740 \ubaa8\ub378\uc758 \uc131\ub2a5\uc5d0 \uc911\ub300\ud55c \uc601\ud5a5\uc744 \ubbf8\uce69\ub2c8\ub2e4. \uc544\ub798\ub294 \uc8fc\uc694 \ud30c\ub77c\ubbf8\ud130\uc785\ub2c8\ub2e4:<\/p>\n
YOLO \ubaa8\ub378\uc758 \uc131\ub2a5\uc740 \uac01\uc885 \ud30c\ub77c\ubbf8\ud130\uc758 \uc870\uc815\uc5d0 \ub530\ub77c \ud06c\uac8c \ub2ec\ub77c\uc9c8 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud559\uc2b5\ub960 \ubc0f \ubc30\uce58 \ud06c\uae30\ub97c \ucd5c\uc801\ud654\ud558\ub294 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\ud559\uc2b5\ub960\uc740 \ubaa8\ub378\uc758 \uc131\ub2a5 \ud5a5\uc0c1\uc5d0 \ub9e4\uc6b0 \uc911\uc694\ud55c \uc5ed\ud560\uc744 \ud569\ub2c8\ub2e4. \ub108\ubb34 \ub192\uc740 \ud559\uc2b5\ub960\uc740 \ubaa8\ub378\uc744 \ubc1c\uc0b0\uc2dc\ud0ac \uc218 \uc788\uc73c\uba70, \ub108\ubb34 \ub0ae\uc740 \ud559\uc2b5\ub960\uc740 \ud6c8\ub828 \uc18d\ub3c4\ub97c \ub290\ub9ac\uac8c \ub9cc\ub4e4\uace0 \uc9c0\uc5ed \ucd5c\uc801\uac12\uc5d0 \uac07\ud790 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
import torch\nimport torchvision\n\n# \ubaa8\ub378 \ucd08\uae30\ud654\nmodel = torchvision.models.yolo()\n\n# \ud559\uc2b5\ub960 \uc9c0\uc815\nlearning_rate = 0.001\nfor param_group in optimizer.param_groups:\n param_group['lr'] = learning_rate\n<\/code><\/pre>\n\ud559\uc2b5\ub960\uc744 \uc870\uc815\ud558\uae30 \uc704\ud574\uc11c\ub294 \ud559\uc2b5\ub960 \uac10\uc18c \uc804\ub7b5\uc744 \ucc44\ud0dd\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud559\uc2b5\ub960 \uac10\uc18c\ub294 \uc5d0\ud3ed \uc218\uac00 \uc99d\uac00\ud568\uc5d0 \ub530\ub77c \ud559\uc2b5\ub960\uc744 \uc904\uc5ec\uc8fc\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4.<\/p>\n
2.2 \ubc30\uce58 \ud06c\uae30 \uc870\uc815<\/h3>\n
\ubc30\uce58 \ud06c\uae30\ub294 GPU \uba54\ubaa8\ub9ac\uc640 \uad00\ub828\uc774 \uc788\uc73c\uba70, \uc801\uc808\ud55c \ud06c\uae30\ub97c \uc124\uc815\ud558\uba74 \ud559\uc2b5 \uc18d\ub3c4\uac00 \ube68\ub77c\uc9c0\uace0 \uc131\ub2a5\uc774 \ud5a5\uc0c1\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uadf8\ub7ec\ub098 \ubc30\uce58 \ud06c\uae30\uac00 \ub108\ubb34 \ud06c\uba74 \uacfc\uc801\ud569\uc5d0 \uc774\ub97c \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
batch_size = 16 # \uae30\ubcf8 \ubc30\uce58 \ud06c\uae30 \uc124\uc815\n\n# DataLoader \ucd08\uae30\ud654\ntrain_loader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=True)\n<\/code><\/pre>\n3. NMS(Non-Maximum Suppression) \uc784\uacc4\uac12 \ucd5c\uc801\ud654<\/h2>\n
YOLO\uc5d0\uc11c\ub294 \uac1d\uccb4 \ud0d0\uc9c0 \ud6c4 NMS\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc911\ubcf5\ub41c \ud0d0\uc9c0\ub97c \uc81c\uac70\ud569\ub2c8\ub2e4. NMS \uc784\uacc4\uac12\uc744 \uc801\uc808\ud558\uac8c \uc870\uc815\ud568\uc73c\ub85c\uc368 \ubaa8\ub378 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
3.1 NMS \uc791\ub3d9 \uc6d0\ub9ac<\/h3>\n
NMS\ub294 \uad50\ucc28 \uc810\uc218\uac00 \uc9c0\uc815\ub41c \uc784\uacc4\uac12\uc744 \ucd08\uacfc\ud560 \uacbd\uc6b0 \uc0c1\uc704 \ubc15\uc2a4\ub97c \uc120\ud0dd\ud558\uc5ec \ub098\uba38\uc9c0\ub97c \uc81c\uac70\ud569\ub2c8\ub2e4. \uc774\ub85c \uc778\ud574 \uc911\ubcf5\ub41c \ud0d0\uc9c0\uac00 \ucd5c\uc18c\ud654\ub429\ub2c8\ub2e4.<\/p>\n
3.2 NMS \uc784\uacc4\uac12 \uc124\uc815<\/h3>\n
NMS \uc784\uacc4\uac12\uc740 \uc77c\ubc18\uc801\uc73c\ub85c 0.3\uc5d0\uc11c 0.7 \uc0ac\uc774\uc758 \uac12\uc73c\ub85c \uc124\uc815\ub429\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \ucd5c\uc801\uc758 \uac12\uc744 \ucc3e\uae30 \uc704\ud574\uc11c\ub294 \uc2e4\ud5d8\uc744 \ud1b5\ud574 \ub9e4\uac1c\ubcc0\uc218\ub97c \uc870\uc815\ud560 \ud544\uc694\uac00 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
def non_max_suppression(predictions, conf_threshold=0.5, nms_threshold=0.3):\n # \uc911\uac04 \ubcc0\uc218 \uc0dd\uc131\n boxes = []\n scores = []\n for prediction in predictions:\n # \uc2e0\ub8b0\ub3c4 \uae30\uc900 \ud544\ud130\ub9c1\n if prediction['confidence'] >= conf_threshold:\n boxes.append(prediction['bbox'])\n scores.append(prediction['confidence'])\n\n # NMS \uc801\uc6a9\n selected_indices = torchvision.ops.nms(boxes, scores, nms_threshold)\n return [boxes[i] for i in selected_indices]\n<\/code><\/pre>\n4. \uc131\ub2a5 \ud5a5\uc0c1 \ubc29\ubc95<\/h2>\n
YOLO\uc758 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud55c \uba87 \uac00\uc9c0 \ubc29\ubc95\uc744 \uc18c\uac1c\ud569\ub2c8\ub2e4. \ub370\uc774\ud130 \uc99d\uac15, \uc804\uc774 \ud559\uc2b5, \ucd5c\uc801\ud654 \uc54c\uace0\ub9ac\uc998\uc744 \ud65c\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
4.1 \ub370\uc774\ud130 \uc99d\uac15<\/h3>\n
\ub370\uc774\ud130 \uc99d\uac15 \uae30\ubc95\uc740 \ubaa8\ub378 \ud6c8\ub828 \uc2dc \ub2e4\uc591\ud55c \ubcc0\ud615\ub41c \ubc84\uc804\uc744 \uacf5\uae09\ud558\uc5ec \ub370\uc774\ud130\uc758 \ub2e4\uc591\uc131\uc744 \ub192\uc774\uace0 \uc624\ubc84\ud53c\ud305\uc744 \ubc29\uc9c0\ud569\ub2c8\ub2e4. \ud68c\uc804, \ud06c\uae30 \ubcc0\uacbd, \uce7c\ub77c \ubcc0\ud658, \ub79c\ub364 \uc790\ub974\uae30 \ub4f1\uc774 \ud3ec\ud568\ub429\ub2c8\ub2e4.<\/p>\n
from torchvision import transforms\n\ndata_transforms = transforms.Compose([\n transforms.RandomHorizontalFlip(),\n transforms.RandomRotation(10),\n transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),\n transforms.Resize((416, 416)),\n transforms.ToTensor(),\n])\n<\/code><\/pre>\n4.2 \uc804\uc774 \ud559\uc2b5<\/h3>\n
\uc804\uc774 \ud559\uc2b5\uc740 \uc0ac\uc804 \ud6c8\ub828\ub41c \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud559\uc2b5\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4. \uc791\uc740 \ub370\uc774\ud130\uc14b\uc5d0\uc11c \uc720\uc6a9\ud558\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc73c\uba70, \ud2b9\ud788 \ube44\uc2b7\ud55c \ub3c4\uba54\uc778\uc5d0\uc11c \ud6a8\uacfc\uc801\uc785\ub2c8\ub2e4.<\/p>\n
from torchvision import models\n\n# \uc0ac\uc804 \ud6c8\ub828\ub41c YOLO \ubaa8\ub378 \ub85c\ub4dc\nmodel = models.detection.yolo(pretrained=True)\n<\/code><\/pre>\n4.3 \ucd5c\uc801\ud654 \uc54c\uace0\ub9ac\uc998<\/h3>\n
\ucd5c\uc801\ud654 \uc54c\uace0\ub9ac\uc998(\uc608: Adam, SGD \ub4f1)\uc744 \ud1b5\ud574 \ud559\uc2b5 \uc131\ub2a5\uc744 \ub354\uc6b1 \ud5a5\uc0c1\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \ubc29\ubc95\uc740 \ubaa8\ub378 \uc218\ub834 \uc18d\ub3c4\ub97c \ub192\uc774\uace0 \uc131\ub2a5\uc744 \uac1c\uc120\ud569\ub2c8\ub2e4.<\/p>\n
import torch.optim as optim\n\n# Adam \ucd5c\uc801\ud654 \uc0ac\uc6a9\noptimizer = optim.Adam(model.parameters(), lr=0.001)\n<\/code><\/pre>\n\uacb0\ub860<\/h2>\n
YOLO \ubaa8\ub378\uc758 \ud30c\ub77c\ubbf8\ud130 \ud29c\ub2dd\uacfc NMS \ucd5c\uc801\ud654\ub294 \ubaa8\ub378 \uc131\ub2a5\uc5d0 \uc911\uc694\ud55c \uc601\ud5a5\uc744 \ubbf8\uce69\ub2c8\ub2e4. \uc704\uc758 \ubc29\ubc95\ub4e4\uc744 \ud1b5\ud574 \ubcf8\uc778\uc758 \ub370\uc774\ud130\uc640 \uc0c1\ud669\uc5d0 \ub9de\uac8c \uac01\uac01\uc758 \ud30c\ub77c\ubbf8\ud130\ub97c \uc870\uc815\ud558\uc5ec \ucd5c\uc801\uc758 \uc131\ub2a5\uc744 \ub2ec\uc131\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc131\ub2a5\uc744 \ub192\uc774\uae30 \uc704\ud574 \uc2e4\ud5d8\uacfc \uac80\uc99d\uc758 \uacfc\uc815\uc744 \ubc18\ubcf5\ud558\ub294 \uac83\uc774 \uc911\uc694\ud558\uba70, \uc774\ub97c \ud1b5\ud574 YOLO \ubaa8\ub378\uc744 \ud65c\uc6a9\ud55c \uac1d\uccb4 \ud0d0\uc9c0\uc758 \uae30\ud68c\ub97c \uadf9\ub300\ud654\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\ubd80\ub85d: \ucd94\uac00 \uc790\ub8cc<\/h2>\n
\ubcf8 \uac15\uc758\uc5d0\uc11c\ub294 YOLO \ubaa8\ub378\uc758 \ud30c\ub77c\ubbf8\ud130 \ud29c\ub2dd\uacfc \ucd5c\uc801\ud654 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \ucd94\uac00\uc801\uc73c\ub85c YOLO\uc5d0 \ub300\ud55c \uacf5\uc2dd \ubb38\uc11c\ub098 \uad00\ub828 \ub17c\ubb38\uc744 \ucc38\uace0\ud558\uae38 \ucd94\ucc9c\ub4dc\ub9bd\ub2c8\ub2e4.<\/p>\n