![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260919050226-6aae176293d22.jpg\")
<\/p>\nYOLOv8-Pose \u662f Ultralytics \u63d0\u4f9b\u7684\u59ff\u6001\u4f30\u8ba1\u6a21\u578b，\u9ed8\u8ba4\u652f\u6301 COCO \u6570\u636e\u96c6\u7684 17 \u4e2a\u5173\u952e\u70b9\u3002\u5bf9\u4e8e\u4e00\u4e9b\u7cbe\u7ec6\u5316\u4efb\u52a1，\u4f8b\u5982\u4eba\u8138 68 \u5173\u952e\u70b9\u68c0\u6d4b\u3001\u624b\u90e8\u5173\u952e\u70b9\u68c0\u6d4b\u6216\u5de5\u4e1a\u573a\u666f\u4e2d\u7684\u5bc6\u96c6\u5173\u952e\u70b9\u56de\u5f52，17 \u4e2a\u70b9\u7684\u8868\u8fbe\u7c92\u5ea6\u5f80\u5f80\u4e0d\u591f\u3002\u901a\u8fc7\u4fee\u6539\u6a21\u578b\u7684 kpt_shape \u4ee5\u53ca\u51c6\u5907\u5bf9\u5e94\u683c\u5f0f\u7684\u6570\u636e\u96c6，\u53ef\u4ee5\u5f88\u65b9\u4fbf\u5730\u628a YOLOv8-Pose \u6269\u5c55\u5230 68 \u4e2a\u5173\u952e\u70b9，\u5728\u4fdd\u6301\u5b9e\u65f6\u6027\u80fd\u7684\u540c\u65f6\u83b7\u5f97\u66f4\u7cbe\u7ec6\u7684\u5750\u6807\u56de\u5f52\u80fd\u529b\u3002<\/p>\n
\u672c\u7bc7\u6559\u7a0b\u5c06\u5b8c\u6574\u8d70\u901a\u300c\u73af\u5883\u642d\u5efa\u300168 \u70b9\u6570\u636e\u51c6\u5907\u3001\u6a21\u578b\u914d\u7f6e\u4fee\u6539\u3001\u8bad\u7ec3\u3001\u9a8c\u8bc1\u3001\u63a8\u7406\u90e8\u7f72\u300d\u5168\u6d41\u7a0b，\u6240\u6709\u793a\u4f8b\u5747\u57fa\u4e8e Ultralytics YOLOv8 \u6846\u67b6\u3002<\/p>\n
\u672c\u7bc7\u6559\u7a0b\u8fd8\u9644\u5e26\u4fee\u6539\u5b66\u4e60\u6570\u636e\u96c6\u5df2\u7ecf\u8f6c\u5316\u5b8c\u6210\u5e76\u514d\u8d39\u5f00\u6e90\u4f7f\u7528，\u6ce8：<\/p>\n
\u3010\u6570\u636e\u96c6\u7b80\u4ecb\u3011 \u672c\u6570\u636e\u96c6\u7531 300W-LP \u6570\u636e\u96c6（AFW \u5b50\u96c6）\u8f6c\u6362\u800c\u6765，\u4e13\u95e8\u7528\u4e8e YOLOv8-Pose \u7684\u4eba\u813868\u5173\u952e\u70b9\u68c0\u6d4b\u8bad\u7ec3\u3002 \u539f\u59cb\u6570\u636e\u6765\u6e90：Face Alignment in Full Pose Range: A 3D Total Solution (TPAMI 2017)\u3002 \u3010\u9884\u5904\u7406\u8bf4\u660e\u3011 \u8f6c\u6362\u65f6\u653e\u5f03\u4e86\u539f\u6570\u636e\u96c6\u4e0e\u56fe\u50cf\u5c3a\u5bf8\u4e0d\u5339\u914d\u7684 roi \u8fb9\u754c\u6846，\u6539\u752868\u4e2a\u6709\u6548\u5173\u952e\u70b9\u6700\u5c0f\u5916\u63a5\u77e9\u5f62\u5e76\u5411\u5916\u6269\u5c5510%\u4f5c\u4e3a\u8fb9\u754c\u6846\u3002 \u5df2\u8f6c\u6362\u4e3a\u6807\u51c6\u7684 YOLO-Pose txt \u683c\u5f0f \u6570\u636e\u96c6\u5df2\u5168\u90e8\u6309 9:1 \u968f\u673a\u5212\u5206\u4e3a\u8bad\u7ec3\u96c6（images\/train, labels\/train）\u548c\u9a8c\u8bc1\u96c6（images\/val, labels\/val）\u3002 \u3010\u9002\u7528\u573a\u666f\u3011 \u4eba\u813868\u5173\u952e\u70b9\u68c0\u6d4b\u3001\u9a7e\u9a76\u5458\u75b2\u52b3\u68c0\u6d4b（EAR\/MAR\u8ba1\u7b97）\u3001YOLOv8-pose \u7b97\u6cd5\u590d\u73b0\u4e0e\u8bfe\u7a0b\u8bbe\u8ba1\/\u6bd5\u4e1a\u8bba\u6587\u5b9e\u9a8c\u3002<\/p>\n
\u3010\u514d\u8d23\u58f0\u660e\u3011 \u7248\u6743\u5f52\u5c5e：\u672c\u6570\u636e\u96c6\u7684\u539f\u59cb\u56fe\u50cf\u53ca\u6807\u6ce8\u7248\u6743\u5f52 300W-LP \u6570\u636e\u96c6\u539f\u4f5c\u8005（Xiangyu Zhu \u7b49）\u53ca\u5bf9\u5e94\u673a\u6784\u6240\u6709\u3002\u672c\u8d44\u6e90\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u3001\u79d1\u7814\u590d\u73b0\u800c\u8fdb\u884c\u7684\u4e8c\u6b21\u683c\u5f0f\u8f6c\u6362\u3002 \u975e\u5546\u4e1a\u7528\u9014：\u672c\u8d44\u6e90\u5b8c\u5168\u514d\u8d39\u5171\u4eab，\u4ec5\u9650\u4e8e\u4e2a\u4eba\u5b66\u4e60\u3001\u5b66\u672f\u7814\u7a76\u7b49\u975e\u5546\u4e1a\u7528\u9014\u3002\u4e25\u7981\u5c06\u672c\u6570\u636e\u96c6\u53ca\u5176\u884d\u751f\u6a21\u578b\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u76c8\u5229\u884c\u4e3a\u3002 \u51c6\u786e\u6027\u58f0\u660e：\u8f6c\u6362\u811a\u672c\u5df2\u5c3d\u53ef\u80fd\u5904\u7406\u5f02\u5e38\u6570\u636e，\u4f46\u53d7\u9650\u4e8e\u539f\u59cb\u6570\u636e\u96c6\u8d28\u91cf，\u65e0\u6cd5\u4fdd\u8bc1\u6240\u670968\u4e2a\u5173\u952e\u70b9\u5747\u7edd\u5bf9\u7cbe\u51c6\u3002\u4f7f\u7528\u8005\u9700\u81ea\u884c\u5bf9\u6570\u636e\u8d28\u91cf\u8fdb\u884c\u6838\u9a8c\u4e0e\u53ef\u89c6\u5316\u62bd\u67e5，\u56e0\u6570\u636e\u7cbe\u5ea6\u5bfc\u81f4\u7684\u4e00\u5207\u540e\u679c\u7531\u4f7f\u7528\u8005\u81ea\u884c\u627f\u62c5\u3002 \u4fb5\u6743\u8054\u7cfb：\u82e5\u672c\u8d44\u6e90\u65e0\u610f\u4e2d\u4fb5\u72af\u4e86\u60a8\u7684\u6743\u76ca，\u8bf7\u901a\u8fc7\u5e73\u53f0\u79c1\u4fe1\u8054\u7cfb\u672c\u4eba，\u6211\u5c06\u5728\u7b2c\u4e00\u65f6\u95f4\u5220\u9664\u8d44\u6e90\u5e76\u81f4\u6b49\u3002 \u4f7f\u7528\u98ce\u9669：\u4e0b\u8f7d\u548c\u4f7f\u7528\u672c\u8d44\u6e90\u5373\u8868\u793a\u60a8\u5df2\u9605\u8bfb\u5e76\u540c\u610f\u672c\u58f0\u660e，\u56e0\u4f7f\u7528\u672c\u8d44\u6e90\u4ea7\u751f\u7684\u4efb\u4f55\u76f4\u63a5\u6216\u95f4\u63a5\u635f\u5931，\u4e0a\u4f20\u8005\u4e0d\u627f\u62c5\u4efb\u4f55\u6cd5\u5f8b\u8d23\u4efb\u3002<\/p>\n
\u5efa\u8bae\u4f7f\u7528 Python 3.9 \u53ca\u4ee5\u4e0a\u7248\u672c\u3002\u521b\u5efa\u4e00\u4e2a\u5e72\u51c0\u7684\u865a\u62df\u73af\u5883，\u5e76\u5b89\u88c5 PyTorch \u548c Ultralytics\u3002<\/p>\n
# \u521b\u5efa\u865a\u62df\u73af\u5883
\npython -m venv yolov8_pose_env
\n# \u6fc0\u6d3b\u865a\u62df\u73af\u5883
\nLinux \/ macOS
\nsource yolov8_pose_env\/bin\/activate
\nWindows
\nyolov8_pose_env\\\\Scripts\\\\activate
\n# \u5b89\u88c5 PyTorch，CUDA \u7248\u672c\u6839\u636e\u5b9e\u9645\u663e\u5361\u73af\u5883\u9009\u62e9
\npip install torch torchvision –index-url https:\/\/download.pytorch.org\/whl\/cu121
\n# \u5b89\u88c5 Ultralytics
\npip install ultralytics <\/p>\n
\u5b89\u88c5\u5b8c\u6210\u540e，\u53ef\u4ee5\u7528\u4e0b\u9762\u7684\u547d\u4ee4\u786e\u8ba4\u73af\u5883\u6b63\u5e38\u3002<\/p>\n
from ultralytics import YOLO
\nimport torch
\nprint("Ultralytics \u7248\u672c：", import("ultralytics").version)
\nprint("CUDA \u662f\u5426\u53ef\u7528：", torch.cuda.is_available()) <\/p>\n
<\/p>\n
68 \u70b9\u6807\u51c6\u987a\u5e8f\u8bf4\u660e<\/p>\n
0\u201316：\u4e0b\u5df4\u8f6e\u5ed3<\/p>\n<\/li>\n
17\u201326：\u7709\u6bdb<\/p>\n<\/li>\n
27\u201335：\u9f3b\u5b50<\/p>\n<\/li>\n
36\u201347：\u773c\u775b<\/p>\n<\/li>\n
48\u201367：\u5634\u5df4<\/p>\n<\/li>\n<\/ul>\n<\/li>\n
68 \u70b9\u7d22\u5f15\u56fe：\u5728\u539f\u56fe\u4e0a\u6807\u51fa\u6bcf\u4e2a\u70b9\u7f16\u53f7\u3002<\/p>\n<\/li>\n
\u5de6\u53f3\u5bf9\u79f0\u70b9\u4ea4\u6362\u8868：\u505a\u00a0fliplr\u00a0\u65f6\u5fc5\u987b\u4ea4\u6362\u54ea\u4e9b\u7d22\u5f15\u3002<\/p>\n<\/li>\n
\u5173\u952e\u70b9\u5750\u6807\u662f\u201c\u76f8\u5bf9\u6574\u56fe\u5f52\u4e00\u5316\u201d，\u4e0d\u662f\u76f8\u5bf9 bbox \u5c40\u90e8\u5f52\u4e00\u5316\u3002<\/p>\n<\/li>\n
\u53ef\u89c1\u6027\u00a0v\u00a0\u7684\u8bed\u4e49：0\u00a0\u4e0d\u53ef\u89c1\u30011\u00a0\u906e\u6321\u30012\u00a0\u53ef\u89c1；\u7f3a\u5931\u70b9\u5982\u4f55\u5904\u7406\u3002<\/p>\n<\/li>\n
\u5355\u8138\/\u591a\u8138\u6807\u7b7e\u5199\u6cd5：\u4e00\u5f20\u56fe\u591a\u4e2a\u4eba\u8138\u65f6，\u6bcf\u884c\u4e00\u4e2a\u76ee\u6807\u3002<\/p>\n<\/li>\n
bbox \u5982\u4f55\u751f\u6210：\u4eba\u8138\u68c0\u6d4b\u5668\u9884\u6807\u6ce8\u3001\u624b\u5de5\u6846\u3001\u6269\u5c55\u6846，\u907f\u514d\u88c1\u6389\u4e0b\u5df4\u6216\u989d\u5934\u3002<\/p>\n<\/li>\n
\u6807\u7b7e\u6821\u9a8c\u811a\u672c：\u68c0\u67e5\u5b57\u6bb5\u6570\u662f\u5426\u4e3a\u00a05 + 68*3\u3001\u5750\u6807\u662f\u5426\u8d8a\u754c\u3001\u53ef\u89c1\u6027\u662f\u5426\u5408\u6cd5\u3001bbox \u662f\u5426\u6709\u6548\u3002<\/p>\n<\/li>\n<\/ul>\n
YOLOv8-Pose \u4f7f\u7528 YOLO \u683c\u5f0f\u7684\u6587\u672c\u6807\u7b7e，\u6bcf\u5f20\u56fe\u7247\u5bf9\u5e94\u4e00\u4e2a\u540c\u540d .txt \u6587\u4ef6\u300268 \u5173\u952e\u70b9\u6807\u7b7e\u7684\u6bcf\u4e00\u884c\u683c\u5f0f\u5982\u4e0b：<\/p>\n
class_id cx cy w h x1 y1 v1 x2 y2 v2 … x68 y68 v68 <\/p>\n
\u5176\u4e2d：<\/p>\n
\u6570\u636e\u96c6\u76ee\u5f55\u5efa\u8bae\u6309\u7167\u4ee5\u4e0b\u7ed3\u6784\u7ec4\u7ec7：<\/p>\n
face68_dataset\/
\n\u251c\u2500\u2500 images\/
\n\u2502 \u251c\u2500\u2500 train\/
\n\u2502 \u2502 \u251c\u2500\u2500 img_001.jpg
\n\u2502 \u2502 \u2514\u2500\u2500 img_002.jpg
\n\u2502 \u2514\u2500\u2500 val\/
\n\u2502 \u251c\u2500\u2500 img_101.jpg
\n\u2502 \u2514\u2500\u2500 img_102.jpg
\n\u2514\u2500\u2500 labels\/
\n \u251c\u2500\u2500 train\/
\n \u2502 \u251c\u2500\u2500 img_001.txt
\n \u2502 \u2514\u2500\u2500 img_002.txt
\n \u2514\u2500\u2500 val\/
\n \u251c\u2500\u2500 img_101.txt
\n \u2514\u2500\u2500 img_102.txt <\/p>\n
\u5982\u679c\u539f\u59cb\u6807\u6ce8\u6765\u81ea\u5e38\u89c1\u7684 68 \u70b9\u6570\u636e\u96c6，\u4f8b\u5982\u4eba\u8138 landmark \u6570\u636e\u96c6，\u901a\u5e38\u9700\u8981\u5148\u628a\u5173\u952e\u70b9\u5750\u6807\u5f52\u4e00\u5316，\u5e76\u8865\u9f50\u53ef\u89c1\u6027\u6807\u7b7e\u3002\u4e0b\u9762\u7684\u793a\u4f8b\u5c55\u793a\u4e86\u5982\u4f55\u628a\u4e00\u7ec4\u5173\u952e\u70b9\u5199\u5165 YOLOv8-Pose \u6807\u7b7e\u6587\u4ef6：<\/p>\n
import os
\nimport cv2
\nimage_dir = "face68_dataset\/images\/train"
\nlabel_dir = "face68_dataset\/labels\/train"
\nos.makedirs(label_dir, exist_ok=True)
\ndef write_yolo_label(image_path, bbox, landmarks68, visibility=None):
\n"""
\nbbox: (x1, y1, x2, y2) \u50cf\u7d20\u5750\u6807
\nlandmarks68: [(x, y), …] \u50cf\u7d20\u5750\u6807，\u5171 68 \u4e2a\u70b9
\nvisibility: \u957f\u5ea6 68 \u7684\u53ef\u89c1\u6027\u5217\u8868，\u9ed8\u8ba4\u5168\u90e8\u4e3a 2
\n"""
\nimg = cv2.imread(image_path)
\nh, w = img.shape[:2]
\nx1, y1, x2, y2 = bbox
\ncx = (x1 + x2) \/ 2 \/ w
\ncy = (y1 + y2) \/ 2 \/ h
\nbw = (x2 – x1) \/ w
\nbh = (y2 – y1) \/ h
\nif visibility is None:
\nvisibility = [2] * 68
\nline_parts = [0, f"{cx:.6f}", f"{cy:.6f}", f"{bw:.6f}", f"{bh:.6f}"]
\nfor (lx, ly), vis in zip(landmarks68, visibility):
\nline_parts.append(f"{lx \/ w:.6f}")
\nline_parts.append(f"{ly \/ h:.6f}")
\nline_parts.append(str(vis))
\nlabel_path = os.path.join(label_dir, os.path.splitext(os.path.basename(image_path))[0] + ".txt")
\nwith open(label_path, "w") as f:
\nf.write(" ".join(line_parts))<\/code><\/pre> <\/p>\n
\u6570\u636e\u51c6\u5907\u597d\u4e4b\u540e，\u8fd8\u9700\u8981\u7f16\u5199\u5bf9\u5e94\u7684\u6570\u636e\u96c6\u914d\u7f6e\u6587\u4ef6 face68.yaml：<\/p>\n
path: \/absolute\/path\/to\/face68_dataset
\ntrain: images\/train
\nval: images\/val
\nnames:
\n0: face
\nkpt_shape: [68, 3] <\/p>\n
Ultralytics \u4ece yolov8n-pose.yaml \u5230 yolov8x-pose.yaml \u63d0\u4f9b\u4e86\u4e0d\u540c\u5c3a\u5bf8\u7684 pose \u6a21\u578b\u914d\u7f6e\u3002\u9ed8\u8ba4\u914d\u7f6e\u4e2d\u7684\u5173\u952e\u70b9\u6570\u91cf\u662f 17，\u6211\u4eec\u9700\u8981\u628a\u5b83\u6539\u6210 68\u3002\u53ef\u4ee5\u76f4\u63a5\u590d\u5236\u5b98\u65b9\u914d\u7f6e\u5e76\u4fee\u6539\u5173\u952e\u70b9\u53c2\u6570\u3002 \u53c2\u8003\u5b98\u65b9 yolov8n-pose.yaml，\u4fee\u6539\u5173\u952e\u70b9\u6570\u91cf\u548c\u5206\u7c7b\u6570<\/p>\n
# yolov8n-face68-pose.yaml
\nnc: 1
\nscales:
\nn: [0.33, 0.25, 1024]
\nbackbone:
\n[-1, 1, Conv, [64, 3, 2]]
\n[-1, 1, Conv, [128, 3, 2]]
\n[-1, 3, C2f, [128, True]]
\n[-1, 1, Conv, [256, 3, 2]]
\n[-1, 6, C2f, [256, True]]
\n[-1, 1, Conv, [512, 3, 2]]
\n[-1, 6, C2f, [512, True]]
\n[-1, 1, Conv, [1024, 3, 2]]
\n[-1, 3, C2f, [1024, True]]
\n[-1, 1, SPPF, [1024, 5]]
\nhead:
\n[-1, 1, nn.Upsample, [None, 2, 'nearest']]
\n[[-1, 6], 1, Concat, [1]]
\n[-1, 3, C2f, [512]]
\n[-1, 1, nn.Upsample, [None, 2, 'nearest']]
\n[[-1, 4], 1, Concat, [1]]
\n[-1, 3, C2f, [256]]
\n[-1, 1, Conv, [256, 3, 2]]
\n[[-1, 15], 1, Concat, [1]]
\n[-1, 3, C2f, [512]]
\n[-1, 1, Conv, [512, 3, 2]]
\n[[-1, 12], 1, Concat, [1]]
\n[-1, 3, C2f, [1024]]
\n[[17, 20, 23], 1, Pose, [nc, kpt_shape]] <\/p>\n
\u6ce8\u610f\u6700\u540e\u4e00\u884c\u4f7f\u7528\u7684\u662f kpt_shape \u53d8\u91cf，\u56e0\u6b64\u53ea\u8981\u5728\u8bad\u7ec3\u65f6\u901a\u8fc7\u53c2\u6570\u628a kpt_shape \u6307\u5b9a\u4e3a [68, 3]，\u6a21\u578b\u8f93\u51fa\u5c42\u7684\u901a\u9053\u6570\u5c31\u4f1a\u81ea\u52a8\u9002\u914d\u4e3a 68 \u4e2a\u5173\u952e\u70b9\u3002<\/p>\n
\u8bad\u7ec3\u811a\u672c\u5f88\u7b80\u5355，\u6838\u5fc3\u662f\u6570\u636e\u96c6\u914d\u7f6e\u548c\u5173\u952e\u70b9\u5f62\u72b6\u3002\u4e0b\u9762\u662f\u4e00\u4e2a\u5b8c\u6574\u8bad\u7ec3\u793a\u4f8b： \u4ece\u5b98\u65b9 pose \u6a21\u578b\u6743\u91cd\u5f00\u59cb\u8bad\u7ec3，\u4e5f\u53ef\u4ee5\u4ece\u81ea\u5b9a\u4e49 yaml \u8bad\u7ec3<\/p>\n
from ultralytics import YOLO<\/p>\n
if __name__ == '__main__':
\n model = YOLO('weights\/yolov8n-pose.pt')<\/p>\n
model.train(data='face.yaml',
\n epochs=300,
\n imgsz=640,
\n batch=2,
\n nbs=16,
\n lr0=0.001,
\n lrf=0.001,
\n warmup_epochs=10,
\n cos_lr=True,
\n optimizer='AdamW',
\n weight_decay=0.0005,
\n mosaic=0.0,
\n mixup=0.0,
\n copy_paste=0.0,
\n fliplr=0.5,
\n flipud=0.0,
\n degrees=10.0,
\n translate=0.05,
\n scale=0.3,
\n shear=0.0,
\n perspective=0.0,
\n hsv_h=0.015,
\n hsv_s=0.5,
\n hsv_v=0.3,
\n pose=8.0,
\n kobj=1.0,
\n cls=0.5,
\n box=7.5,
\n dfl=1.5,
\n workers=0,
\n amp=True,
\n patience=100,
\n save_period=10,
\n project='runs\/pose',
\n name='face68_nano',
\n cache=False,
\n ) <\/p>\n
\u8fd9\u91cc\u6709\u51e0\u4e2a\u9700\u8981\u7279\u522b\u6ce8\u610f\u7684\u53c2\u6570：<\/p>\n
mosaic=0.0：\u5173\u95edMosaic \u589e\u5f3a\u8fd9\u662f\u5bf9 68 \u70b9\u4eba\u8138\u5173\u952e\u70b9\u6700\u6709\u5bb3\u7684\u589e\u5f3a\u3002\u62fc\u56fe\u4f1a\u628a\u4eba\u8138\u5207\u788e\u3001\u7f29\u653e\u3001\u9519\u4f4d，\u5173\u952e\u70b9\u5750\u6807\u968f\u4e4b\u88ab\u6253\u4e71\u3002\u6a21\u578b\u8981\u82b1\u5927\u91cf\u7cbe\u529b\u201c\u89e3\u62fc\u56fe\u201d，\u800c\u4e0d\u662f\u5b66\u4eba\u8138\u7ed3\u6784\u3002 kpt_shape=[68, 3]：\u5fc5\u987b\u662f\u4e09\u7ef4\u5217\u8868\u3002\u7b2c\u4e09\u7ef4\u901a\u5e38\u662f\u5173\u952e\u70b9\u7684\u53ef\u89c1\u6027\u6807\u5fd7；\u5982\u679c\u6807\u6ce8\u53ea\u6709 x\u3001y \u800c\u6ca1\u6709\u53ef\u89c1\u6027，\u53ef\u4ee5\u5199 [68, 2]\u3002 imgsz：\u4eba\u8138\u6216\u5c0f\u76ee\u6807\u573a\u666f\u53ef\u4ee5\u9002\u5f53\u589e\u5927\u5230 640 \u6216 1024，\u4f46\u663e\u5b58\u5360\u7528\u4e5f\u4f1a\u589e\u52a0\u3002 batch：\u6839\u636e\u663e\u5b58\u8c03\u6574，\u663e\u5b58\u4e0d\u8db3\u65f6\u51cf\u5c0f batch \u5e76\u9002\u5f53\u964d\u4f4e\u5b66\u4e60\u7387\u3002 patience：\u65e9\u505c\u8f6e\u6570，\u5efa\u8bae\u6839\u636e\u6570\u636e\u96c6\u89c4\u6a21\u8bbe\u5b9a\u3002 \u8bad\u7ec3\u8fc7\u7a0b\u4e2d Ultralytics \u4f1a\u81ea\u52a8\u8bb0\u5f55 loss\u3001mAP \u548c\u5404\u7c7b\u53ef\u89c6\u5316\u56fe\u50cf，\u7ed3\u679c\u9ed8\u8ba4\u4fdd\u5b58\u5728 runs\/pose\/yolov8n_face68 \u76ee\u5f55\u4e0b\u3002<\/p>\n
\u8bad\u7ec3\u5b8c\u6210\u540e，\u4f7f\u7528\u9a8c\u8bc1\u96c6\u8bc4\u4f30\u6a21\u578b\u6548\u679c：<\/p>\n
from ultralytics import YOLO
\nmodel = YOLO("runs\/pose\/yolov8n_face68\/weights\/best.pt")y <\/p>\n
\u5728\u9a8c\u8bc1\u96c6\u4e0a\u8bc4\u4f30<\/p>\n
metrics = model.val(data="face68.yaml", kpt_shape=[68, 3])
\nprint(metrics.box.map)
\nprint(metrics.box.map50)
\nprint(metrics.box.map75) <\/p>\n
\u5bf9\u5355\u5f20\u56fe\u7247\u6216\u89c6\u9891\u505a\u63a8\u7406\u65f6，\u53ef\u4ee5\u4f7f\u7528\u4e0b\u9762\u7684\u65b9\u6cd5：<\/p>\n
from ultralytics import YOLO
\nmodel = YOLO("runs\/pose\/yolov8n_face68\/weights\/best.pt")
\nresults = model.predict(
\nsource="test_faces.jpg",
\nkpt_shape=[68, 3],
\nconf=0.25,
\nsave=True,
\nshow=True,
\n)
\nfor result in results:
\nkeypoints = result.keypoints
\nboxes = result.boxes
\nif keypoints is not None:
\nkeypoints.xy：\u5173\u952e\u70b9\u5750\u6807，\u5f62\u72b6\u4e3a (\u76ee\u6807\u6570, 68, 2)
\nkeypoints.conf：\u5173\u952e\u70b9\u7f6e\u4fe1\u5ea6，\u5f62\u72b6\u4e3a (\u76ee\u6807\u6570, 68)
\nprint("\u5173\u952e\u70b9\u5750\u6807：", keypoints.xy)
\nprint("\u5173\u952e\u70b9\u7f6e\u4fe1\u5ea6：", keypoints.conf) <\/p>\n
\u5982\u679c\u9700\u8981\u628a\u5173\u952e\u70b9\u7ed3\u679c\u5bfc\u51fa\u4e3a JSON，\u53ef\u4ee5\u7ed3\u5408 result.keypoints \u624b\u52a8\u7ec4\u7ec7\u6570\u636e\u7ed3\u6784\u3002\u9ed8\u8ba4\u60c5\u51b5\u4e0b，result.keypoints.xy \u662f 68 \u4e2a\u5f52\u4e00\u5316\u6216\u50cf\u7d20\u683c\u5f0f\u7684\u5750\u6807，\u5177\u4f53\u53d6\u51b3\u4e8e\u63a8\u7406\u8bbe\u7f6e\u3002<\/p>\n
\u8bad\u7ec3\u5b8c\u6210\u540e，\u53ef\u4ee5\u628a best.pt \u5bfc\u51fa\u4e3a ONNX\u3001TensorRT \u7b49\u683c\u5f0f\u3002\u65e0\u8bba\u5bfc\u51fa\u54ea\u79cd\u683c\u5f0f，\u90fd\u8981\u4f20\u5165\u4e0e\u8bad\u7ec3\u65f6\u4e00\u81f4\u7684 kpt_shape=[68, 3]，\u5426\u5219 Ultralytics \u65e0\u6cd5\u6b63\u786e\u786e\u5b9a\u5173\u952e\u70b9\u8f93\u51fa\u901a\u9053\u6570，\u5bfc\u51fa\u7684\u6a21\u578b\u53ef\u80fd\u4ecd\u6309\u9ed8\u8ba4 17 \u70b9\u89e3\u6790\u3002<\/p>\n
\u4e0b\u9762\u5148\u5bfc\u51fa ONNX：<\/p>\n
from ultralytics import YOLO
\nmodel = YOLO("runs\/pose\/yolov8n_face68\/weights\/best.pt")
\nmodel.export(
\nformat="onnx", # \u5bfc\u51fa ONNX \u683c\u5f0f
\nimgsz=640, # \u4f7f\u7528\u4e0e\u8bad\u7ec3\u65f6\u4e00\u81f4\u7684\u8f93\u5165\u5c3a\u5bf8
\nkpt_shape=[68, 3], # \u5173\u952e\u70b9\u6570\u91cf\u548c\u53ef\u89c1\u6027\u7ef4\u5ea6
\nopset=12, # ONNX opset \u7248\u672c
\nsimplify=True, # \u7b80\u5316\u6a21\u578b\u56fe，\u4fbf\u4e8e\u90e8\u7f72
\n) <\/p>\n
\u5bfc\u51fa TensorRT engine \u65f6\u540c\u7406，\u901a\u5e38\u5728 NVIDIA GPU \u4e0a\u4f7f\u7528 FP16 \u52a0\u901f：<\/p>\n
model.export(
\n format="engine", # \u5bfc\u51fa TensorRT \u683c\u5f0f
\n imgsz=640,
\n kpt_shape=[68, 3],
\n half=True, # \u5f00\u542f\u534a\u7cbe\u5ea6\u63a8\u7406，\u9700 GPU \u652f\u6301
\n device=0,
\n) <\/p>\n
\u5bfc\u51fa\u7684 ONNX \u8f93\u51fa\u5f20\u91cf\u53ef\u4ee5\u7406\u89e3\u4e3a [1, 4 + nc + 68 * 3, num_anchors]\u3002\u524d 4 \u4e2a\u901a\u9053\u662f\u76ee\u6807\u6846\u56de\u5f52\u503c，\u63a5\u7740\u662f nc \u4e2a\u7c7b\u522b\u901a\u9053，\u6700\u540e 204 \u4e2a\u901a\u9053\u6309 68 \u7ec4\u3001\u6bcf\u7ec4 3 \u4e2a\u503c\u7ec4\u7ec7\u4e3a x\u3001y\u3001\u53ef\u89c1\u6027\u3002\u4e0b\u9762\u7528 ONNX Runtime \u8bfb\u53d6\u8f93\u51fa\u5e76\u62c6\u5206\u5173\u952e\u70b9\u901a\u9053：<\/p>\n
import numpy as np
\nimport onnxruntime as ort
\nsession = ort.InferenceSession("runs\/pose\/yolov8n_face68\/weights\/best.onnx")
\ninput_name = session.get_inputs()[0].name
\ninput_tensor \u9700\u8981\u9884\u5148\u5b8c\u6210\u7f29\u653e\u3001\u5f52\u4e00\u5316\u548c BGR \u5230 RGB \u7684\u8f6c\u6362
\noutput = session.run(None, {input_name: input_tensor})[0]
\nprint("\u539f\u59cb\u8f93\u51fa\u5f62\u72b6：", output.shape) # [1, 4 + nc + 68*3, num_anchors]
\nnc = 1
\nnum_anchors = output.shape[2]
\npred = output[0].transpose(1, 0) # [num_anchors, 4 + nc + 68*3]
\nboxes_raw = pred[:, :4] # \u76ee\u6807\u6846\u56de\u5f52\u539f\u59cb\u503c
\ncls_raw = pred[:, 4:4 + nc] # \u7c7b\u522b\u539f\u59cb\u503c
\nkpts_raw = pred[:, 4 + nc:].reshape(num_anchors, 68, 3)
\nkpts_xy = kpts_raw[…, :2] # 68 \u4e2a\u5173\u952e\u70b9\u7684 x\u3001y \u539f\u59cb\u503c
\nkpts_conf = kpts_raw[…, 2] # \u5173\u952e\u70b9\u53ef\u89c1\u6027\u6216\u7f6e\u4fe1\u5ea6
\nprint("boxes_raw \u5f62\u72b6：", boxes_raw.shape)
\nprint("kpts_xy \u5f62\u72b6：", kpts_xy.shape)
\nprint("kpts_conf \u5f62\u72b6：", kpts_conf.shape) <\/p>\n
\u9700\u8981\u7279\u522b\u6ce8\u610f，kpts_xy \u548c boxes_raw \u4ecd\u662f\u6a21\u578b\u8f93\u51fa\u7684\u539f\u59cb\u56de\u5f52\u503c，\u4e0d\u662f\u6700\u7ec8\u50cf\u7d20\u5750\u6807\u3002\u5b9e\u9645\u90e8\u7f72\u65f6\u8fd8\u9700\u8981\u7ed3\u5408 anchor\u3001stride \u5b8c\u6210\u89e3\u7801，\u5e76\u8fdb\u884c NMS\u3002\u5bf9\u4e8e\u5927\u591a\u6570\u9879\u76ee，\u63a8\u8350\u76f4\u63a5\u901a\u8fc7 Ultralytics \u52a0\u8f7d ONNX \u6a21\u578b\u63a8\u7406，\u5b83\u80fd\u81ea\u52a8\u5b8c\u6210\u89e3\u7801\u548c\u540e\u5904\u7406，\u5e76\u76f4\u63a5\u8fd4\u56de 68 \u70b9\u5750\u6807：<\/p>\n
from ultralytics import YOLO
\nmodel = YOLO("runs\/pose\/yolov8n_face68\/weights\/best.onnx")
\nresults = model.predict(
\nsource="test_faces.jpg",
\nkpt_shape=[68, 3],
\nconf=0.25,
\n)
\nfor result in results:
\nif result.keypoints is not None:
\nxy = result.keypoints.xy # \u5f62\u72b6\u4e3a [\u76ee\u6807\u6570, 68, 2]
\nconf = result.keypoints.conf # \u5f62\u72b6\u4e3a [\u76ee\u6807\u6570, 68]
\nprint("68 \u5173\u952e\u70b9\u5750\u6807：", xy)
\nprint("\u5173\u952e\u70b9\u7f6e\u4fe1\u5ea6：", conf) <\/p>\n
\u5982\u679c kpt_shape \u4e0e\u6807\u7b7e\u6587\u4ef6\u4e2d\u7684\u5173\u952e\u70b9\u6570\u91cf\u4e0d\u4e00\u81f4，\u8bad\u7ec3\u542f\u52a8\u65f6\u901a\u5e38\u4f1a\u51fa\u73b0\u7ef4\u5ea6\u9519\u8bef\u3002\u8bf7\u68c0\u67e5\u6570\u636e\u96c6\u914d\u7f6e\u4e2d\u7684 kpt_shape: [68, 3] \u4ee5\u53ca\u8bad\u7ec3\u547d\u4ee4\u4e2d\u662f\u5426\u4e5f\u90fd\u4f20\u5165\u4e86 kpt_shape=[68, 3]\u3002<\/p>\n
68 \u4e2a\u5173\u952e\u70b9\u6bd4 17 \u4e2a\u5173\u952e\u70b9\u7684\u8f93\u51fa\u901a\u9053\u591a，\u4f46\u5bf9\u9aa8\u5e72\u7f51\u7edc\u7684\u663e\u5b58\u5f71\u54cd\u8fdc\u5c0f\u4e8e\u8f93\u5165\u5206\u8fa8\u7387\u548c batch \u5e26\u6765\u7684\u5f71\u54cd\u3002\u663e\u5b58\u4e0d\u8db3\u65f6\u4f18\u5148\u51cf\u5c0f imgsz \u6216 batch，\u5176\u6b21\u662f\u6362\u66f4\u8f7b\u91cf\u7684 yolov8n-pose \u4e3b\u5e72\u3002<\/p>\n
\u901a\u5e38\u4e0e\u5173\u952e\u70b9\u5750\u6807\u5f52\u4e00\u5316\u9519\u8bef\u6709\u5173\u3002\u8bf7\u786e\u8ba4\u6240\u6709 x\u3001y \u5750\u6807\u90fd\u9664\u4ee5\u4e86\u56fe\u7247\u5bbd\u9ad8，\u4e14\u4e0d\u662f\u9664\u4ee5 1 \u5230 68 \u7684\u7d22\u5f15\u3002\u53e6\u5916\u68c0\u67e5\u53ef\u89c1\u6027\u5b57\u6bb5\u662f\u5426\u88ab\u8bef\u5f53\u4f5c\u5750\u6807\u5199\u5165\u3002<\/p>\n
\u4e0d\u5efa\u8bae\u76f4\u63a5\u6df7\u7528\u300268 \u70b9\u6a21\u578b\u548c 17 \u70b9\u6a21\u578b\u8f93\u51fa\u5c42\u7ed3\u6784\u4e0d\u540c，\u4e0d\u80fd\u76f4\u63a5\u52a0\u8f7d\u5f7c\u6b64\u7684\u5173\u952e\u70b9\u5934\u90e8\u6743\u91cd\u3002\u5982\u679c\u5fc5\u987b\u590d\u7528，\u53ef\u4ee5\u5c06\u9aa8\u5e72\u7f51\u7edc\u6743\u91cd\u8fc1\u79fb\u540e\u91cd\u65b0\u8bad\u7ec3\u5173\u952e\u70b9\u5934\u90e8\u3002<\/p>\n
\u5f53\u6807\u6ce8\u89e3\u6790\u6216\u6570\u636e\u589e\u5f3a\u8fc7\u7a0b\u4e2d\u5750\u6807\u53d8\u6362\u4e0d\u4e00\u81f4\u65f6，68 \u5173\u952e\u70b9\u7ecf\u5e38\u4f1a\u51fa\u73b0\u6574\u4f53\u504f\u79fb\u3001\u5de6\u53f3\u955c\u50cf\u6216\u5c3a\u5ea6\u5f02\u5e38\u3002\u4e0b\u8868\u6c47\u603b\u4e86\u4e09\u79cd\u5178\u578b\u75c7\u72b6\u7684\u6392\u67e5\u65b9\u5411\u3002<\/p>\n
| \u5750\u6807\u6574\u4f53\u504f\u79fb<\/td>\n | \u88c1\u526a\u3001\u4eff\u5c04\u53d8\u6362\u6216\u7f29\u653e\u540e\u6ca1\u6709\u540c\u6b65\u8c03\u6574\u5173\u952e\u70b9；\u5173\u952e\u70b9\u76f8\u5bf9\u6574\u56fe\u548c\u76f8\u5bf9 bbox \u4f7f\u7528\u4e86\u4e0d\u540c\u7684\u539f\u70b9\u3002<\/td>\n | \u786e\u4fdd\u6240\u6709\u70b9\u5728\u540c\u4e00\u4e2a\u53d8\u6362\u77e9\u9635\u4e0b\u5904\u7406；\u68c0\u67e5\u5173\u952e\u70b9\u5750\u6807\u662f\u76f8\u5bf9\u6574\u5f20\u56fe\u5f52\u4e00\u5316，\u800c\u4e0d\u662f\u76f8\u5bf9 bbox \u5c40\u90e8\u533a\u57df\u3002<\/td>\n<\/tr>\n |
| \u5de6\u53f3\u955c\u50cf<\/td>\n | \u6c34\u5e73\u7ffb\u8f6c\u56fe\u50cf\u540e\u6ca1\u6709\u540c\u6b65\u7ffb\u8f6c\u5173\u952e\u70b9；\u6216\u8005\u53ea\u7ffb\u8f6c\u4e86\u56fe\u7247，\u6ca1\u6709\u4ea4\u6362 68 \u70b9\u4e2d\u5de6\u53f3\u5bf9\u79f0\u7684\u7d22\u5f15\u3002<\/td>\n | \u5bf9\u56fe\u50cf\u505a\u6c34\u5e73\u7ffb\u8f6c\u65f6，\u540c\u6b65\u628a\u5173\u952e\u70b9 x \u5750\u6807\u6539\u4e3a 1-x，\u5e76\u6309 68 \u70b9\u5bf9\u79f0\u5173\u7cfb\u4ea4\u6362\u5de6\u53f3\u70b9\u7d22\u5f15，\u540c\u65f6\u66f4\u65b0\u53ef\u89c1\u6027\u3002<\/td>\n<\/tr>\n |
| \u5c3a\u5ea6\u5f02\u5e38<\/td>\n | \u7f29\u653e\u524d\u540e\u6ca1\u6709\u4f7f\u7528\u540c\u4e00\u7ec4\u56fe\u7247\u5bbd\u9ad8\u8fdb\u884c\u5f52\u4e00\u5316，\u6216\u628a\u50cf\u7d20\u5750\u6807\u4e0e\u5f52\u4e00\u5316\u5750\u6807\u6df7\u7528\u3002<\/td>\n | \u7edf\u4e00\u9664\u4ee5\u539f\u56fe\u5bbd\u9ad8；\u5bf9\u7f29\u653e\u540e\u7684\u5750\u6807\u5148\u8fd8\u539f\u5230\u539f\u56fe\u518d\u9a8c\u8bc1；\u589e\u5f3a\u65f6\u8bb0\u5f55\u5e76\u590d\u7528\u5b9e\u9645\u7684\u7f29\u653e\u548c padding \u53c2\u6570\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u4e0b\u9762\u8fd9\u6bb5\u8c03\u8bd5\u4ee3\u7801\u53ef\u4ee5\u8bfb\u53d6\u4e00\u5f20\u56fe\u7247\u53ca\u5176 YOLO \u683c\u5f0f\u7684 68 \u70b9\u6807\u7b7e，\u5c06\u5173\u952e\u70b9\u7ed8\u5236\u56de\u539f\u56fe，\u4fbf\u4e8e\u4eba\u5de5\u68c0\u67e5\u5750\u6807\u504f\u79fb\u3001\u955c\u50cf\u5173\u7cfb\u548c\u5c3a\u5ea6\u662f\u5426\u6b63\u5e38\u3002<\/p>\n import cv2 \u8fd0\u884c\u540e\u91cd\u70b9\u68c0\u67e5\u5173\u952e\u70b9\u662f\u5426\u843d\u5728\u4eba\u8138\u3001\u624b\u90e8\u6216\u5176\u4ed6\u76ee\u6807\u7ed3\u6784\u4e0a，\u5c24\u5176\u5173\u6ce8\u5de6\u53f3\u5bf9\u79f0\u70b9\u548c\u6574\u4f53\u8f6e\u5ed3\u5c3a\u5ea6\u3002\u82e5\u53d1\u73b0\u6574\u4f53\u504f\u79fb\u3001\u955c\u50cf\u6216\u7f29\u653e\u5f02\u5e38，\u5e94\u56de\u5230\u6807\u6ce8\u89e3\u6790\u548c\u6570\u636e\u589e\u5f3a\u73af\u8282\u8fdb\u884c\u4fee\u6b63\u3002<\/p>\n 8. \u603b\u7ed3<\/h3>\n\u5c06 YOLOv8-Pose \u6269\u5c55\u5230 68 \u5173\u952e\u70b9\u7684\u5173\u952e\u6b65\u9aa4\u53ef\u4ee5\u5f52\u7eb3\u4e3a：\u51c6\u5907 68 \u70b9 YOLO \u683c\u5f0f\u6807\u7b7e\u3001\u5728\u6570\u636e\u914d\u7f6e\u4e2d\u58f0\u660e kpt_shape: [68, 3]\u3001\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u65f6\u663e\u5f0f\u4f20\u5165\u76f8\u540c\u7684 kpt_shape，\u5e76\u901a\u8fc7\u5c0f\u89c4\u6a21\u6570\u636e\u5148\u8dd1\u901a\u6d41\u7a0b\u518d\u6269\u91cf\u8bad\u7ec3\u3002\u638c\u63e1\u8fd9\u4e00\u5957\u6d41\u7a0b\u540e，\u53ef\u4ee5\u8fdb\u4e00\u6b65\u8fc1\u79fb\u5230\u66f4\u591a\u70b9\u7684\u624b\u90e8\u3001\u59ff\u6001\u6216\u5b9a\u5236\u5316\u5173\u952e\u70b9\u4efb\u52a1\u3002 \u5efa\u8bae\u5728\u6b63\u5f0f\u8bad\u7ec3\u524d，\u5148\u7528\u51e0\u767e\u5f20\u6837\u672c\u9a8c\u8bc1\u6570\u636e\u683c\u5f0f\u548c\u6a21\u578b\u8f93\u51fa\u7ef4\u5ea6\u662f\u5426\u6b63\u786e，\u518d\u9010\u6b65\u589e\u52a0\u6570\u636e\u89c4\u6a21\u548c\u8bad\u7ec3\u8f6e\u6570，\u8fd9\u6837\u80fd\u66f4\u9ad8\u6548\u5730\u6392\u67e5\u5173\u952e\u70b9\u56de\u5f52\u4e0d\u7a33\u5b9a\u7684\u95ee\u9898\u3002<\/p>\n \u52a0\u6570\u636e\u89c4\u6a21\u548c\u8bad\u7ec3\u8f6e\u6570，\u8fd9\u6837\u80fd\u66f4\u9ad8\u6548\u5730\u6392\u67e5\u5173\u952e\u70b9\u56de\u5f52\u4e0d\u7a33\u5b9a\u7684\u95ee\u9898\u3002<\/p>\n \u53c2\u8003\u8d44\u6599<\/h3>\nUltralytics YOLOv8 \u5b98\u65b9\u6587\u6863：https:\/\/docs.ultralytics.com\/models\/yolov8\/ Ultralytics YOLOv8 Pose \u4efb\u52a1\u6587\u6863：https:\/\/docs.ultralytics.com\/tasks\/pose\/ 300W-LP \u6570\u636e\u96c6\u539f\u59cb\u8bba\u6587：Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, Stan Z. Li. Face Alignment in Full Pose Range: A 3D Total Solution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. YOLOv8-Pose \u76f8\u5173 GitHub \u4ed3\u5e93：https:\/\/github.com\/ultralytics\/ultralytics ONNX Runtime \u90e8\u7f72\u6587\u6863：https:\/\/onnxruntime.ai\/docs\/ TensorRT \u90e8\u7f72\u6587\u6863：https:\/\/docs.nvidia.com\/deeplearning\/tensorrt\/<\/p><\/p>\n","protected":false},"excerpt":{"rendered":" 1. \u4e3a\u4ec0\u4e48\u9009\u62e9 YOLOv8-Pose \u8bad\u7ec3 68 \u5173\u952e\u70b9 |