はじめる

Ultralytics YOLOによる姿勢推定

Ultralytics YOLO pose estimation with human body keypoint detection

姿勢推定は、画像内の特定の点(一般にキーポイントと呼ばれます)の位置を特定するタスクです。キーポイントは、関節、ランドマーク、その他の特徴的な部分など、オブジェクトのさまざまな部位を表します。キーポイントの位置は通常、2D [x, y] 座標の集合で表され、任意で可視性フラグ [x, y, visible] を付けることができます。

姿勢推定モデルの出力は、画像内のオブジェクトのキーポイントを表す点のセットで、通常は各点の信頼度スコアも含まれます。シーン内のオブジェクトの特定の部分と、それら相互の位置関係を特定する必要がある場合、姿勢推定が適しています。



視聴: Ultralytics YOLO26 でリアルタイム姿勢推定を実行する方法 | トラッキングとキーポイント抽出 🕺
ヒント

YOLO26 Poseモデルでは -pose サフィックスを使用します(例:yolo26n-pose.pt)。COCOキーポイントで事前学習されています。

モデル#

COCOキーポイントデータセットで事前学習されたYOLO26 Poseモデルを以下に示します。

モデルは、初回使用時に最新のUltralytics リリースから自動的にダウンロードされます。

モデルサイズ
(ピクセル)
mAPpose50-95
(e2e)
mAPpose50
(e2e)
速度
CPU ONNX
(ms)
速度
T4 TensorRT10
(ms)
パラメーター
(M)
FLOPs
(B)
YOLO26n-pose64057.283.340.3 ± 0.51.8 ± 0.02.97.6
YOLO26s-pose64063.086.685.3 ± 0.92.7 ± 0.010.424.1
YOLO26m-pose64068.889.6218.0 ± 1.55.0 ± 0.121.573.3
YOLO26l-pose64070.490.5275.4 ± 2.46.5 ± 0.125.991.7
YOLO26x-pose64071.691.6565.4 ± 3.012.2 ± 0.257.6202.3
  • mAPvalの値は、COCO Keypoints val2017データセットで、単一モデル・単一スケールを使用した場合のものです。
    yolo pose val data=coco-pose.yaml device=0 nms=Falseで再現できます
  • 速度は、COCO val画像に対してONNXをCPUで、TensorRT10をNVIDIA T4 GPUで使用した場合の平均値です。
    yolo pose val data=coco-pose.yaml batch=1 device=0|cpu nms=Falseで再現できます。
  • ParamsとFLOPsの値は、Conv/BatchNormのfoldingと未使用の検出ブランチの削除を適用した後の、融合済みモデルの値です。事前学習済みチェックポイントにはトレーニング時のアーキテクチャ全体が保持されているため、より大きな値が表示される場合があります。

予備的なCOCOキーポイントの結果については、未リリースのYOLO27プレビューをご覧ください。

学習#

COCO8-poseデータセットで、画像サイズ640、100エポックの条件でYOLO26n-poseモデルをトレーニングします。

例
from ultralytics import YOLO

# モデルを読み込む
model = YOLO("yolo26n-pose.yaml")  # YAMLから新しいモデルを構築します
model = YOLO("yolo26n-pose.pt")  # 事前トレーニング済みモデルを読み込みます(トレーニングにはこちらを推奨します)
model = YOLO("yolo26n-pose.yaml").load("yolo26n-pose.pt")  # YAMLからモデルを構築し、重みを転送します

# モデルをトレーニングする
results = model.train(data="coco8-pose.yaml", epochs=100, imgsz=640)

trainモードの詳細については、学習ページをご覧ください。姿勢推定モデルは、Ultralytics Platformのクラウドトレーニングでも学習できます。

データセットのフォーマット#

YOLO姿勢データセット形式の詳細については、データセットガイドをご覧ください。既存のCOCOキーポイントJSONデータセットをYOLO形式に変換するには、COCOからYOLOへのガイドで説明されているように、組み込みのconvert_coco関数をuse_keypoints=Trueとともに使用します。Ultralytics Platformのアノテーションでは、人、手、顔、およびカスタムのキーポイント配置に対応した組み込みのスケルトンテンプレートを使って、姿勢ラベルを付けることもできます。

カスタム姿勢推定タスクでは、動物の姿勢推定に使えるTiger-Pose、手のトラッキングに使えるHand Keypoints、犬の姿勢分析に使えるDog-Poseなどの専門データセットもご検討ください。

検証#

トレーニング済みYOLO26n-poseモデルの精度を、平均適合率(mAP)として測定して検証します。model はトレーニング時の data と引数をモデル属性として保持するため、引数を指定する必要はありません。トレーニングの例にある path/to/best.pt はCOCO8-poseで検証します。公式の重みには、使用環境に存在しないトレーニングデータセットのパスが記録されているため、警告を表示してタスクのデフォルト coco8-pose.yaml にフォールバックします。別のデータセットで検証するには、data を渡してください。

例
from ultralytics import YOLO

# モデルを読み込む
model = YOLO("yolo26n-pose.pt")  # 公式モデルを読み込みます
model = YOLO("path/to/best.pt")  # カスタムモデルを読み込む

# モデルを検証します
metrics = model.val()  # データセットと設定が記憶されるため、引数は不要です
metrics.box.map  # map50-95
metrics.box.map50  # map50
metrics.box.map75  # map75
metrics.box.maps  # クラスごとの mAP50-95。検証セットに存在しないクラスには全体の mAP が設定されます
metrics.box.ap_class_index  # 実際に評価されたクラスのインデックス
metrics.box.image_metrics  # ボックスの画像ごとのメトリクス辞書。precision、recall、F1、TP、FP、FNを含みます
metrics.pose.map  # map50-95(P)
metrics.pose.map50  # map50(P)
metrics.pose.map75  # map75(P)
metrics.pose.maps  # クラスごとの mAP50-95(P)。検証セットに存在しないクラスには全体の mAP(P) が設定されます
metrics.pose.ap_class_index  # 実際に評価されたクラスのインデックス
metrics.pose.image_metrics  # 姿勢の画像ごとのメトリクス辞書。precision、recall、F1、TP、FP、FNを含みます

推論#

学習済みのYOLO26n-poseモデルを使用して、画像に対する予測を実行します。

例
from ultralytics import YOLO

# モデルを読み込む
model = YOLO("yolo26n-pose.pt")  # 公式モデルを読み込みます
model = YOLO("path/to/best.pt")  # カスタムモデルを読み込む

# モデルで予測します
results = model("https://ultralytics.com/images/bus.jpg")  # 画像に対して予測します

# 結果にアクセスします
for result in results:
    xy = result.keypoints.xy  # x座標とy座標
    xyn = result.keypoints.xyn  # normalized
    kpts = result.keypoints.data  # x、y、可視性(利用可能な場合)

Predictページでpredictモードの詳細をご確認ください。

結果の出力#

姿勢推定では、画像ごとに1つのResultsオブジェクトが返されます。主な予測フィールドは、姿勢の座標を示すresult.keypointsと、それらのキーポイントが属する検出インスタンスを示すresult.boxesです。

属性種類形状説明
result.keypointsKeypoints(N)キーポイントです。
result.keypoints.datatorch.float32(N,K,2/3)x,yに、オプションで可視性/信頼度が加わります。
result.keypoints.xytorch.float32(N,K,2)ピクセル座標のキーポイントです。
result.keypoints.xyntorch.float32(N,K,2)正規化されたキーポイントです。
result.boxesBoxes(N)インスタンスボックスです。

すべてのタスクにおける、タスク固有のResultsフィールドについては、タスク別の予測結果セクションをご覧ください。

キーポイントのインデックスマップ#

デフォルトのYOLO26 Poseモデルは17個のCOCOキーポイントを予測するため、result.keypoints.xy[:, k]内のインデックスkは次のとおりです。

  1. 鼻
  2. 左目
  3. 右目
  4. 左耳
  5. 右耳
  6. 左肩
  7. 右肩
  8. 左肘
  9. 右肘
  10. 左手首
  11. 右手首
  12. 左腰
  13. 右腰
  14. 左膝
  15. 右膝
  16. 左足首
  17. 右足首

エクスポート#

YOLO26n-poseモデルをONNX、CoreMLなどの別の形式にエクスポートします。

例
from ultralytics import YOLO

# モデルを読み込む
model = YOLO("yolo26n-pose.pt")  # 公式モデルを読み込みます
model = YOLO("path/to/best.pt")  # カスタムモデルを読み込む

# モデルをエクスポートする
model.export(format="onnx")

使用可能なYOLO26-poseのエクスポート形式を以下の表に示します。format 引数を使用して、任意の形式にエクスポートできます(例:format='onnx' または format='engine')。エクスポート済みモデルに対して直接予測または検証を実行できます(例:yolo predict model=yolo26n-pose.onnx)。エクスポートが完了すると、モデルの使用例が表示されます。

形式format引数モデルメタデータ引数
PyTorch-yolo26n-pose.pt✅-
TorchScripttorchscriptyolo26n-pose.torchscript✅imgsz, quantize, dynamic, nms, batch, device
ONNXonnxyolo26n-pose.onnx✅imgsz, quantize, dynamic, simplify, opset, nms, batch, data, fraction, device
OpenVINOopenvinoyolo26n-pose_openvino_model/✅imgsz, quantize, dynamic, nms, batch, data, fraction, device
TensorRTengineyolo26n-pose.engine✅imgsz, quantize, dynamic, simplify, opset, workspace, nms, batch, data, fraction, device
CoreMLcoremlyolo26n-pose.mlpackage✅imgsz, dynamic, quantize, nms, batch, device
Apple Core AIcoreaiyolo26n-pose.aimodel✅imgsz, batch, quantize
TF SavedModelsaved_modelyolo26n-pose_saved_model/✅imgsz, quantize, opset, nms, batch, data, fraction, device
TF GraphDefpbyolo26n-pose.pb❌imgsz, opset, batch, device
TF Edge TPUedgetpuyolo26n-pose_edgetpu.tflite✅imgsz, quantize, opset, data, fraction, device
LiteRTlitertyolo26n-pose.tflite✅imgsz, quantize, batch, data, fraction, device
PaddlePaddlepaddleyolo26n-pose_paddle_model/✅imgsz, batch, device
MNNmnnyolo26n-pose.mnn✅imgsz, batch, dynamic, quantize, simplify, opset, nms, device
NCNNncnnyolo26n-pose_ncnn_model/✅imgsz, quantize, device
IMX500imxyolo11n_imx_model/✅imgsz, quantize, data, fraction, nms, device
RKNNrknnyolo26n-pose_rknn_model/✅imgsz, batch, name, quantize, simplify, opset, data, fraction, device
ExecuTorchexecutorchyolo26n-pose_executorch_model/✅imgsz, quantize, batch, data, fraction, device
Axeleraaxelerayolo26n-pose_axelera_model/✅imgsz, batch, quantize, data, fraction, device
DEEPXdeepxyolo26n-pose_deepx_model/✅imgsz, quantize, simplify, opset, data, optimize, device
Qualcomm QNNqnnyolo26n-pose_qnn.onnx✅imgsz, batch, name, quantize, simplify, opset, data, fraction, device
Hailohailoyolo26n-pose_hailo_model/✅imgsz, name, quantize, data, fraction, simplify, conf, iou, device
Huawei Ascendascendyolo26n-pose_ascend_model/✅imgsz, batch, name, quantize, opset, simplify, nms, device
AMD Xilinxxilinxyolo26n-pose_xilinx_model/✅imgsz, name, quantize, data, fraction, opset, simplify, device
NXPnxpyolo26n-pose_nxp.tflite✅imgsz, quantize, data, fraction, device
ST Edge AIstedgeaiyolo26n-pose_stedgeai.onnx✅imgsz, quantize, data, fraction, opset, simplify, device

nms=None は、外部NMS用にデフォルトでraw出力を返します(YOLO26はAxelera上でNMSフリーヘッドを維持します)。利用可能なNMSフリーヘッドを選択するには、nms=Falseを設定します。非対応の形式では、ネイティブの出力経路にフォールバックします。上記のnmsの項目は、nms=Trueを使用してNMSを埋め込める形式を示しています。

詳細なexport情報については、Exportページをご覧ください。

よくある質問#

  • Ultralytics YOLO26による姿勢推定では、画像内のキーポイントと呼ばれる特定の点を特定します。キーポイントは通常、関節やオブジェクトのその他の重要な特徴を表します。出力には各点の [x, y] 座標と信頼度スコアが含まれます。YOLO26-poseモデルはこのタスク専用に設計されており、-pose サフィックスを使用します(例:yolo26n-pose.pt)。これらのモデルはCOCOキーポイントなどのデータセットで事前学習されており、さまざまな姿勢推定タスクに使用できます。デフォルトモデルはCOCOの17個のキーポイントを予測します。

  • カスタムデータセットでYOLO26-poseモデルを学習するには、YAMLファイルで定義した新しいモデルまたは事前学習済みモデルを読み込みます。その後、指定したデータセットとパラメーターを使用して学習を開始できます。

    例
    from ultralytics import YOLO
    
    # モデルを読み込む
    model = YOLO("yolo26n-pose.yaml")  # YAMLから新しいモデルを構築します
    model = YOLO("yolo26n-pose.pt")  # 事前トレーニング済みモデルを読み込みます(トレーニングにはこちらを推奨します)
    
    # モデルをトレーニングする
    results = model.train(data="your-dataset.yaml", epochs=100, imgsz=640)

    学習の詳細については、学習セクションをご覧ください。カスタム姿勢推定モデルをノーコードで学習するには、Ultralytics Platformのクラウドトレーニングも利用できます。

  • YOLO26-poseモデルの検証では、学習時に保持されたものと同じデータセットパラメーターを使用して精度を評価します。次に例を示します。

    例
    from ultralytics import YOLO
    
    # モデルを読み込む
    model = YOLO("yolo26n-pose.pt")  # 公式モデルを読み込みます
    model = YOLO("path/to/best.pt")  # カスタムモデルを読み込む
    
    # モデルを検証します
    metrics = model.val()  # データセットと設定が記憶されるため、引数は不要です

    詳細については、検証セクションをご覧ください。

  • はい、YOLO26-poseモデルはONNX、CoreML、TensorRTなどのさまざまな形式にエクスポートできます。Pythonまたはコマンドラインインターフェース(CLI)を使用してエクスポートできます。

    例
    from ultralytics import YOLO
    
    # モデルを読み込む
    model = YOLO("yolo26n-pose.pt")  # 公式モデルを読み込みます
    model = YOLO("path/to/best.pt")  # カスタムモデルを読み込む
    
    # モデルをエクスポートする
    model.export(format="onnx")

    詳細については、エクスポートセクションをご覧ください。エクスポートしたモデルは、フィットネストラッキング、スポーツ分析、ロボティクスなどのリアルタイムアプリケーション向けにエッジデバイスへデプロイできます。

  • Ultralytics YOLO26には、YOLO26n-pose、YOLO26s-pose、YOLO26m-poseなど、さまざまな事前学習済み姿勢モデルがあります。これらのモデルはサイズ、精度(mAP)、速度が異なります。各モデルのCOCOキーポイントmAP、速度、パラメーター数、FLOPsを含む一覧は、モデルセクションをご覧ください。

コメント465

質問したり、結果を共有したり、ほかの人をサポートしたりできます。

  1. 3300299472GitHub

    import pygetwindow as gw import pyautogui from ultralytics import YOLO import cv2

    Load the model

    model = YOLO("model/best.onnx")

    while True: # Get the handle of the application window app_window = gw.getWindowsWithTitle('Chess Game')[0] app_window.activate()

    # Move the window to the top-left corner of the screen
    app_window.moveTo(0, 0)
    
    # Set the window size to (929, 721)
    app_window.resizeTo(929, 721)
    
    # Get the window position
    left, top, width, height = app_window.left, app_window.top, app_window.width, app_window.height
    
    # Take a screenshot in the background
    screenshot = pyautogui.screenshot(region=(left, top, width, height))
    screenshot.save('screenshot.png')
    
    # Read the screenshot and convert it to OpenCV format
    img = cv2.imread('screenshot.png')
    
    # Perform inference
    results = model.predict(source=img, show=True)

    I need help, who will help me? I have successfully identified the object in the program through the ONNX model, but I don't know how to obtain coordinates. I need to obtain coordinates and then control the mouse to move to the coordinates for clicking operations. Can someone show me the reference code? Thank you

    1
    1. Glenn JocherUltralytics(編集済み)

      @3300299472 hello! It's great to hear that you've successfully used the ONNX model for object detection. To obtain the coordinates of the detected objects and control the mouse accordingly, you can extract the bounding box coordinates from the results object and use them to move the mouse. Here's a quick example on how you can do this:

      from ultralytics import YOLO
      import pyautogui
      
      # Load the model
      model = YOLO("model/best.onnx")
      
      # Perform inference
      results = model.predict(source=img, show=True)
      
      # Get bounding box coordinates
      for x1, y1, x2, y2 in results[0].boxes.xyxy.tolist():  # detections in the first image
          # Calculate the center of the bounding box
          x_center = int((x1 + x2) / 2)
          y_center = int((y1 + y2) / 2)
          
          # Move mouse to the center of the bounding box
          pyautogui.moveTo(x_center, y_center)
          # Optionally, perform a click
          pyautogui.click()

      This code snippet will move the mouse to the center of each detected object and perform a click. Make sure to adjust the coordinates based on your specific screen resolution and application window size if necessary. Happy coding! 😊

  2. AdadAlShababGitHub

    Hi. I want to measure the length of detected line after running the pose detection model. How can I do it? Please let me know.

    1
    1. Glenn JocherUltralytics

      Hi! To measure the length of a detected line after running the pose detection model, you can compute the Euclidean distance between keypoints. If you have keypoints' coordinates, you can use the following Python code example:

      import numpy as np
      
      def calculate_distance(pt1, pt2):
          return np.sqrt((pt2[0] - pt1[0])**2 + (pt2[1] - pt1[1])**2)
      
      # Example keypoints coordinates
      keypoint1 = [x1, y1]  # Replace x1, y1 with your actual coordinates
      keypoint2 = [x2, y2]  # Replace x2, y2 with your actual coordinates
      
      # Calculate distance
      length = calculate_distance(keypoint1, keypoint2)
      print('The length of the line is:', length)

      Just replace x1, y1, x2, y2 with the actual coordinates of your keypoints. This will give you the length between these two points which can represent a line in your use case. Hope this helps! 😊

  3. Pakorn-02GitHub(編集済み)

    Can i only detect the first person entering the camera? And from the source code, can I correctly convert from key point coordinates (x, y) to position coordinates (x, y, z) in Unity?

    #My Source Code
    import requests
    import cv2
    import numpy as np
    from ultralytics import YOLO
    
    model = YOLO('yolov8n-pose.pt')
    cap = cv2.VideoCapture(0)
    
    server_url = 'http://localhost:3000/keypoints'
    
    while True:
        success, frame = cap.read()
        results = model(frame)
        
        # Simulate depth data (replace this with actual depth estimation)
        depth_data = np.random.rand(frame.shape[0], frame.shape[1])
        
        # Access the xy attribute and convert it to a list
        keypoints = results[0].keypoints.xy.cpu().numpy().tolist()
        # data = {'keypoints': keypoints} 
        
        #Extract only the keypoints for the right hand
        right_hand_keypoints = [keypoints[0][10]]
        
        # Combine 2D keypoints with simulated depth data to obtain 3D keypoints
        keypoints_3d = [(x, y, depth_data[int(y), int(x)]) for x, y in right_hand_keypoints]
        
        data = {'keypoints': keypoints_3d}
    
        # Send keypoint data to the Node.js server
        response = requests.post(server_url, json=data)
        print(response.text)  # Print server response
    
        annotated_image = results[0].plot()
        cv2.imshow('frame', annotated_image)
        cv2.waitKey(1)
    
    cap.release()
    cv2.destroyAllWindows()
    1
    1. Paula DerrengerUltralytics

      Hi there! Regarding your question about detecting only the first person entering the camera, YOLOv8 models by default will detect all instances of persons in each frame. However, you can modify the behavior to track the first detected person by managing the IDs of detected persons and only processing the first ID observed across frames.

      As for converting from 2D keypoint coordinates (x, y) to 3D position coordinates (x, y, z) in Unity, the conversion depends on getting a real depth measurement. Your approach of simulating depth data and combining it with the keypoints is a common technique. If you have access to a depth sensor or can derive depth from stereo cameras, that would provide more accurate 3D coordinates. Just ensure that the coordinate system alignment between the AI model output and Unity's coordinate system is correctly calibrated for accurate placements in Unity.

      Here's a refined approach considering depth measurements:

      # Assuming depth_data contains actual depth measurements
      keypoints_3d = [(x, y, depth_data[int(y), int(x)]) for x, y in right_hand_keypoints]

      Keep refining your method to handle actual depth, and ensure coordinate transformations align with Unity's requirements. Good luck! 🚀

  4. Pakorn-02GitHub(編集済み)

    How can i convert 2D Key point value (x, y) to Position (x, y, z) on each bone in the body (3D Model) in Unity? Thanks

    1
    1. Glenn JocherUltralytics

      Converting 2D keypoints to a 3D model position in Unity requires some form of depth estimation since 2D keypoints captured from images or video only provide information on the x and y axes. To add the z-axis (depth), you'll typically need additional data or a model that can infer depth from the given information.

      While Ultralytics YOLOv8 specializes in pose estimation providing 2D [x, y] keypoints, integrating these with a 3D model in Unity to achieve the desired [x, y, z] positioning might involve:

      1. Using stereo vision techniques if you have images from multiple angles.
      2. Applying a separate depth estimation model that can predict the z-coordinate based on the 2D image.
      3. Utilizing additional sensors data (like Lidar or structured light sensors) that provides depth information.

      Once you have the depth information, you can map the 2D keypoints [x, y] from YOLOv8 to 3D [x, y, z] coordinates in Unity by aligning the keypoints with the depth information.

      Here's a conceptual snippet for understanding:

      # Example for converting (x, y) keypoints to (x, y, z)
      # This assumes you have a method to retrieve 'z' (depth)
      depth_data = get_depth_information() # Custom method to obtain depth
      keypoints_2d = [(x, y) for x, y in detected_keypoints]  # from YOLOv8
      keypoints_3d = [(x, y, depth_data[x, y]) for x, y in keypoints_2d]

      In Unity, you would then apply these 3D coordinates to position the joints or bones of your model accordingly. Note that exact implementation details would depend on how you're structuring your Unity project and the source of your depth data.

      Remember, working with 3D models and real-world positioning introduces complexities, especially concerning coordinate systems alignment and scale adjustments across different frameworks and data sources. 🎮🕵️‍♂️

    1. Paula DerrengerUltralytics

      @leonidasavila hey there! 👋 To calculate the angle between three points, you can use the law of cosines. Here's a quick example in Python:

      import numpy as np
      
      def calculate_angle(p1, p2, p3):
          # p1, p2, p3 are the points in format [x, y]
          # Calculate the vectors
          v1 = np.array(p1) - np.array(p2)
          v2 = np.array(p3) - np.array(p2)
          
          # Calculate the angle in radians
          angle_rad = np.arccos(np.dot(v1, v2) / (np.linalg.norm(v1) * np.linalg.norm(v2)))
          
          # Convert to degrees
          angle_deg = np.degrees(angle_rad)
          
          return angle_deg
      
      # Example points
      p1 = [0, 0]
      p2 = [1, 0]
      p3 = [1, 1]
      
      angle = calculate_angle(p1, p2, p3)
      print(f"The angle is {angle} degrees")

      This calculates the angle at p2 formed by the line segments p1p2 and p2p3. Hope this helps! Let me know if you have any other questions. 😊

      1
    2. leonidasavilaGitHub

      Thanks @pderrenger, for your quickly response! I´m trying to migrate our startup app from mediapipe to yolo 😍

      Do you have any tutorial or short example to define specifics landmarks to make calculates? (4 example push-up)

      Best regards! Leonidas

    3. Paula DerrengerUltralytics

      Hey Leonidas! 🌟

      It's awesome to hear about your transition from MediaPipe to YOLO for your startup app! For specific landmark detection like push-ups, you can leverage our YOLOv8 pose models. Here's a quick example to get you started:

      from ultralytics import YOLO
      
      # Load a pretrained pose model
      model = YOLO('yolov8n-pose.pt')
      
      # Predict keypoints on an image
      results = model('path/to/your/image.jpg')
      
      # Access keypoints
      keypoints = results[0].keypoints
      
      # Now, you can use these keypoints to calculate angles or any other metrics you need

      This snippet loads a pretrained YOLOv8 pose model and predicts keypoints on your image. You can then use these keypoints for further calculations, like angles between joints during push-ups.

      For more detailed guidance, check out our Pose Estimation page. If you have any more questions or need further assistance, feel free to ask. Happy coding! 😊

    4. ozankavcu38GitHub

      By the way sir after loading the model how could we dissapear the pose estimatet lines like boxes=False argument

    5. Paula DerrengerUltralytics(編集済み)

      Hey there! 😊 To hide the pose estimation lines (similar to using boxes=False for bounding boxes), you can simply not plot them. When you're working with the results from the pose estimation model, you have control over what gets visualized. Here's a quick example on how to display the keypoints without drawing the connecting lines:

      from ultralytics import YOLO
      
      # Load your pose model
      model = YOLO('yolov8n-pose.pt')
      
      # Run prediction
      results = model('path/to/image.jpg')
      
      # Access keypoints
      keypoints = results[0].keypoints
      
      # Now, you can visualize keypoints without drawing lines between them
      # This part depends on how you're visualizing. For example, using OpenCV to plot keypoints:
      import cv2
      image = cv2.imread('path/to/image.jpg')
      for person in keypoints.xy:  # one (17, 2) tensor per detected person
          for x, y in person.tolist():
              cv2.circle(image, (int(x), int(y)), radius=3, color=(0, 255, 0), thickness=-1)
      
      cv2.imshow("Keypoints without lines", image)
      cv2.waitKey(0)
      cv2.destroyAllWindows()

      This way, you'll see the keypoints on your image without the estimated pose lines. Hope this helps, and good luck with your app migration! If you have any more questions, feel free to ask. 🚀

      1
    6. KindstepGitHub

      I am currently using Yolo11 for my thesis in software engineering. Regarding the calculation of angles using keypoints, what is the correct way to access the detected keypoints directly in the code? assuming the indexing the same as in:

      0: Nose 1: Left Eye 2: Right Eye 3: Left Ear 4: Right Ear 5: Left Shoulder 6: Right Shoulder 7: Left Elbow 8: Right Elbow 9: Left Wrist 10: Right Wrist 11: Left Hip 12: Right Hip 13: Left Knee 14: Right Knee 15: Left Ankle 16: Right Ankle

      So assuming i use a similar function as you proposed, to calculate the angle of the right elbow joint a live feed (e.g. webcam) and the code is something like this:

      model = YOLO('yolo11x-pose.pt')

      results = model(frame, show=True, conf=conf, device=device, save=True)

      keypoints = results[0].keypoints

      angle = calculate_angle(keypoints6, keypoints8, keypoints10) # print(f"The angle is {angle} degrees")

      What is the correct way to access each keypoint? Is it indexed in a array like keypoints[6] or what would be the correct way to go forward?

      I hope my question is not too messy. It's getting quite late, but I am enjopying my time working with Yolo11.

      1
    7. Glenn JocherUltralytics

      To access specific keypoints in YOLO11 pose estimation, you can index the keypoints array directly using the indices corresponding to each body part. For example, to calculate the angle at the right elbow, you can use keypoints[6], keypoints[8], and keypoints[10] for the right shoulder, right elbow, and right wrist, respectively. Let me know if you need further assistance!

    8. KindstepGitHub

      Okay, i Already figured it out now that I had time to try it. Seems I was on the right track. :)

      Have a great weekend everybody!

      1
    9. leonidasavilaGitHub

      @kindstep thanks for sharing. Good luck with your thesis, it's will be fine! Currently, we are using YOLO8 and we are studying to migrate YOLO11... I hope the deploy of hand pose very soon!

  5. CosophiaGitHub

    What task does line annotation correspond to? what I know is : rectangle -> detection point -> pose polygon -> Segment

    1. Paula DerrengerUltralytics

      A line annotation does not map to a standalone Ultralytics task. Use pose estimation when the line represents connections between labeled keypoints, or segmentation when you need the line’s pixel-level shape; please continue dropping questions here so the Ultralytics team can assist.

    1. Glenn JocherUltralytics

      Yes — Ultralytics YOLO26 pose models can be used with ByteTrack through Track mode; feel free to post any follow-up details here so the Ultralytics team can see them, and the pose docs and Track mode docs cover the supported workflow.

  6. JC3GitHub

    Ok... but where is the documentation on how to actually run the model and access the results? Seems to be missing.

    1. Glenn JocherUltralytics

      You can run a pose model via Python or CLI; the returned results objects expose keypoints directly—see the usage pattern on the Pose Estimation docs (and feel free to drop any follow-up questions here so the Ultralytics team can see them).

      from ultralytics import YOLO
      
      model = YOLO("yolo26n-pose.pt")
      results = model("path/to/image.jpg")  # or a video path / 0 for webcam
      
      r = results[0]
      xy = r.keypoints.xy     # (n, k, 2) pixel coords
      xyn = r.keypoints.xyn   # (n, k, 2) normalized coords
      kpts = r.keypoints.data # (n, k, 3) x, y, visible/conf (if available)
      yolo pose predict model=yolo26n-pose.pt source=path/to/image.jpg
  7. shayan-mesdaghiGitHub

    Hello everyone, I trained a YOLO model on a dataset I created myself. The training went well, but I have a major problem: why isn't the skeleton (landmarks connections) displayed on my images and videos, and only the keypoints are shown? I performed the labelling process using Roboflow. I should mention that when I use a pre-trained YOLO model, the skeleton is displayed, but when I train the dataset myself, only the keypoints are present. Please guide me, thank you.

    1. Glenn JocherUltralytics

      @shayan-mesdaghi skeleton lines are auto-drawn only for the COCO 17‑keypoint human schema; custom keypoint counts/orders show dots unless you render your own connections. If you’re not using the COCO 17 order, set your skeleton at draw time like this:

      from ultralytics import YOLO
      from ultralytics.utils.plotting import Annotator
      
      # 1-based COCO-17 pairs; adjust to your schema as needed
      skeleton = [(16,14),(14,12),(17,15),(15,13),(12,13),(6,12),(7,13),(6,7),
                  (6,8),(7,9),(8,10),(9,11),(2,3),(1,2),(1,3),(2,4),(3,5),(4,6),(5,7)]
      
      model = YOLO("best.pt")
      for r in model("image_or_video.mp4"):
          im = r.orig_img.copy()
          ann = Annotator(im)
          ann.skeleton = skeleton  # 1-based indices
          for k in r.keypoints.xy:
              ann.kpts(k, kpt_line=True)
          im_out = ann.result()

      If you edited plotting.py, restart your Python session/notebook so changes load. Method details are in the Annotator.kpts reference, which notes the human‑only default for kpt_line see Annotator.kpts, and the dataset YAML fields for pose (e.g., kpt_shape) are shown in the Pose datasets overview.

  8. mobin3399GitHub

    Hi, I’ve trained a custom YOLOv8-pose model with two object classes. One class occupies more than 80% of the image area (large object), while the other one is much smaller. Interestingly, the smaller object produces much more accurate keypoint predictions, while the large object shows poor keypoint localization — even though both classes have precise annotations, balanced samples, and consistent lighting/camera setup.

    I have already tried:

    • Careful labeling (pixel-level accurate keypoints)
    • Strong data augmentation
    • Adjusting loss weights for keypoints and bounding boxes
    • Different model scales (YOLOv8n-pose, YOLOv8m-pose)

    However, the issue remains: pose accuracy drops significantly for the larger object, despite good detection and validation metrics.

    From my analysis, this might be related to:

    1. The OKS-based keypoint loss being scale-dependent (large objects are penalized less for the same absolute error).
    2. The receptive field or feature resolution in the pose head being insufficient for very large objects.
    3. Bounding box coverage (close to image borders) affecting normalized keypoint coordinates.

    Has anyone else observed this behavior in YOLOv8-pose or YOLOv9-pose? And is there a recommended way to make keypoint accuracy more scale-invariant (for example, by modifying the OKS loss or normalization strategy)?

    Any guidance or similar experiences would be greatly appreciated!

    1. pratikbariyaGitHub

      Ok, you have done proper annotation, augmentation, loss weight adjustments, and tried different models. See, there is an issue regarding YOLOV8-POSE for large objects. I think you have to focus on the KPT loss also. If it is low, then try to improve that by making changes in loss (by overriding or code changes).

      If possible, share an image of the object and model matrix. That will help understand the issue.

      1
    2. Glenn JocherUltralytics(編集済み)

      @pratikbariya @mobin3399 adding a few concrete levers you can try for the “large object → poorer keypoint localization” case: train/val at higher resolution and lighter augments for pose, e.g. imgsz=1280 mosaic=0.0 degrees=0 translate=0.05 scale=0.5 fliplr=0.5, and prefer images where the large object isn’t hard‑clipped by borders (pad or pre-crop so it fills the frame); if you’re not using the COCO 17‑kpt schema, note our evaluator defaults to uniform OKS sigmas which can bias big objects—set explicit sigmas for your keypoints or adopt the COCO set; details and usage are in the Pose docs and PoseValidator notes.

      yolo pose train model=yolo26x-pose.pt data=your.yaml imgsz=1280 mosaic=0.0 degrees=0 translate=0.05 scale=0.5 fliplr=0.5

      See the Pose Estimation guide for end‑to‑end training tips and keypoint ordering, and the PoseValidator reference for how OKS is computed. Pose Estimation – Ultralytics YOLO • PoseValidator reference (OKS/sigma behavior)

      1
    3. mobin3399GitHub

      Thank you for your response. I followed your suggestion, and it helped me resolve the issue to a large extent. I really appreciate your guidance and support.