Erste Schritte

Datensatz mit Handschlüsselpunkten

Einführung#

Der Ultralytics-Hand-Keypoints-Datensatz umfasst 26.768 Bilder von Händen, die jeweils mit 21 Keypoints annotiert sind. Sie wurden mithilfe der Google MediaPipe-Bibliothek erstellt, um eine hohe Genauigkeit und Konsistenz zu gewährleisten. Der Datensatz ist mit den Formaten von Ultralytics YOLO26 kompatibel und eignet sich zum Trainieren von Modellen zur Posenschätzung.



Ansehen: Schätzung von Hand-Schlüsselpunkten mit Ultralytics YOLO | Tutorial zur Posenschätzung menschlicher Hände

Schlüsselpunkte#

Diagramm der Handschlüsselpunkte mit 21 Punkten

Jede Hand ist mit den folgenden 21 Schlüsselpunkten annotiert:

  1. Handgelenk
  2. Daumen (4 Punkte)
  3. Zeigefinger (4 Punkte)
  4. Mittelfinger (4 Punkte)
  5. Ringfinger (4 Punkte)
  6. Kleiner Finger (4 Punkte)

Datensatzstruktur#

  • Bilder insgesamt: 26.768 (18.776 train / 7.992 val).
  • Klassen: 1 (hand).
  • Schlüsselpunkte: 21 pro Hand mit Tripeln aus (x, y, visibility).
  • Downloadgröße: ~369 MB.

Für ein eigenes Gestikvokabular, das über allgemeine Handlandmarken hinausgeht, ermöglicht die Ultralytics Platform, deinen eigenen Datensatz direkt im Browser zu annotieren und zu trainieren.

Anwendungsfälle#

Hand-Keypoints unterstützen verschiedene reale Anwendungsfälle:

  • Gestenerkennung: Mensch-Computer-Interaktion und berührungslose Steuerungsschnittstellen.
  • AR/VR-Steuerung: präzise Interaktion mit virtuellen Objekten.
  • Robotermanipulation: präzise Steuerung von Roboterhänden.
  • Gesundheitswesen: Analyse von Handbewegungen für medizinische Diagnosen.
  • Animation: Bewegungserfassung für realistische Handbewegungen.
  • Biometrische Authentifizierung: Sicherheitssysteme auf Grundlage der Handgeometrie.

YAML-Datei des Datensatzes#

Eine YAML-Datei dient zum Festlegen der Datensatzkonfiguration. Sie enthält Informationen zu den Pfaden, Klassen und weiteren relevanten Angaben des Datensatzes. Für den Datensatz mit Handschlüsselpunkten wird die Datei hand-keypoints.yaml unter https://github.com/ultralytics/ultralytics/blob/main/ultralytics/cfg/datasets/hand-keypoints.yaml gepflegt.

ultralytics/cfg/datasets/hand-keypoints.yaml
# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license

# Hand Keypoints dataset by Ultralytics
# Documentation: https://docs.ultralytics.com/datasets/pose/hand-keypoints
# Example usage: yolo train data=hand-keypoints.yaml
# parent
# ├── ultralytics
# └── datasets
#     └── hand-keypoints ← downloads here (369 MB)

# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]
path: hand-keypoints # dataset root dir
train: images/train # train images (relative to 'path') 18776 images
val: images/val # val images (relative to 'path') 7992 images

# Keypoints
kpt_shape: [21, 3] # number of keypoints, number of dims (2 for x,y or 3 for x,y,visible)
flip_idx: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20]

# Classes
names:
  0: hand

# Keypoint names per class
kpt_names:
  0:
    - wrist
    - thumb_cmc
    - thumb_mcp
    - thumb_ip
    - thumb_tip
    - index_mcp
    - index_pip
    - index_dip
    - index_tip
    - middle_mcp
    - middle_pip
    - middle_dip
    - middle_tip
    - ring_mcp
    - ring_pip
    - ring_dip
    - ring_tip
    - pinky_mcp
    - pinky_pip
    - pinky_dip
    - pinky_tip

# Download script/URL (optional)
download: https://github.com/ultralytics/assets/releases/download/v0.0.0/hand-keypoints.zip

Verwendung#

Um ein YOLO26n-pose-Modell für 100 Epochen mit einer Bildgröße von 640 auf dem Datensatz Hand Keypoints zu trainieren, kannst du die folgenden Codebeispiele verwenden.

Trainingsbeispiel
from ultralytics import YOLO

# Ein Modell laden
model = YOLO("yolo26n-pose.pt")  # ein vortrainiertes Modell laden (für das Training empfohlen)

# Das Modell trainieren
results = model.train(data="hand-keypoints.yaml", epochs=100, imgsz=640)

Beispielbilder und Annotationen#

Der Datensatz mit Handschlüsselpunkten enthält vielfältige Bilder menschlicher Hände, die mit Schlüsselpunkten annotiert sind. Hier sind einige Beispielbilder aus dem Datensatz mit den zugehörigen Annotationen:

Beispiel aus dem Datensatz zur Pose-Schätzung mit Handschlüsselpunkten

  • Mosaikbild: Dieses Bild zeigt einen Trainingsbatch aus mosaikartig kombinierten Datensatzbildern. Beim Mosaikverfahren werden während des Trainings mehrere Bilder zu einem einzigen Bild kombiniert, um die Vielfalt der Objekte und Szenen in jedem Trainingsbatch zu erhöhen. Dadurch kann das Modell besser auf unterschiedliche Objektgrößen, Seitenverhältnisse und Kontexte generalisieren.

Das Beispiel veranschaulicht die Vielfalt und Komplexität der Bilder im Datensatz mit Handschlüsselpunkten sowie die Vorteile des Mosaikverfahrens während des Trainings.

Zitate und Danksagungen#

Wenn du den Datensatz mit Handschlüsselpunkten in deiner Forschung oder Entwicklungsarbeit verwendest, nenne bitte die folgenden Quellen:

Danksagungen

Wir möchten den folgenden Quellen für die Bereitstellung der in diesem Datensatz verwendeten Bilder danken:

Die Bilder wurden gemäß den von den jeweiligen Plattformen bereitgestellten Lizenzen gesammelt und verwendet und werden unter der Creative-Commons-Lizenz Namensnennung-Nicht kommerziell-Weitergabe unter gleichen Bedingungen 4.0 International verbreitet.

Wir möchten auch dem Ersteller dieses Datensatzes, Rion Dsilva, dafür danken, dass er ihn der Forschungscommunity für Vision-KI zur Verfügung gestellt hat.

Häufig gestellte Fragen#

  • Lade yolo26n-pose.pt und rufe model.train(data="hand-keypoints.yaml", epochs=100, imgsz=640) auf – im obigen Trainingsbeispiel findest du die vollständigen Python- und CLI-Codeausschnitte. Eine umfassende Liste der Argumente enthält die Seite Training des Modells.

  • Mit 26.768 annotierten Bildern und 21 Keypoints pro Hand, die mithilfe von Google MediaPipe erzeugt wurden, bietet der Datensatz Hand Keypoints Posenschätzungsmodellen den Umfang und die Annotationsgenauigkeit, die für anspruchsvolle Aufgaben der Posenschätzung erforderlich sind.

  • Hand Keypoints unterstützt Gestenerkennung, AR/VR-Steuerung, robotische Manipulation, Bewegungsanalysen im Gesundheitswesen, Animation und biometrische Authentifizierung – im Abschnitt Anwendungen findest du Details zu den einzelnen Bereichen.

  • Der Datensatz Hand Keypoints ist in zwei Teilmengen unterteilt:

    1. Train: Enthält 18.776 Bilder zum Trainieren von Pose-Estimation-Modellen.
    2. Val: Enthält 7.992 Bilder zur Validierung während des Modelltrainings.
  • Die Datensatzkonfiguration ist in einer YAML-Datei definiert, die Pfade, Klassen und weitere relevante Informationen enthält. Die Datei hand-keypoints.yaml findest du unter hand-keypoints.yaml.

    Um diese YAML-Datei für das Training zu verwenden, gib sie wie im obigen Trainingsbeispiel in deinem Trainingsskript oder CLI-Befehl an. Weitere Informationen findest du im Abschnitt Datensatz-YAML.

Kommentare32

Stell eine Frage, teile deine Ergebnisse oder hilf jemand anderem.

  1. MengBo-GitHubGitHub

    感谢如此详细的教学文档,可惜在个人学习使用阶段没有足够的 训练条件 ,我请求分享一些训练后的模型提供给学习者使用,感激不尽😄

    1
    1. Glenn JocherUltralytics(bearbeitet)

      感谢支持!目前我们未提供基于 Hand Keypoints 的官方预训练权重;建议从 YOLO26n‑pose.pt 做迁移学习,或使用 Ultralytics Platform 进行一键云端训练获取可用模型,如有社区训练的 checkpoint 也欢迎在此贴共享链接。

    1. zlj127GitHub

      Wow! Thank you so much for your contribution, bro. It has saved me most of the time for pre-training. You're my god!

      1
  2. RedAbrahamGitHub

    what's the meaning of output(130069)? I wanna deploy the model by myself but I don't know how to analyze the output

    1. Glenn JocherUltralytics(bearbeitet)

      That (1, 300, 69) is the exported end-to-end pose head output shaped as (batch, num_predictions, num_channels); for Hand Keypoints kpt_shape: [21, 3], so 69 = 4 (box x1,y1,x2,y2) + 1 (confidence score) + 1 (class index, names: {0: hand}) + 21*3 (keypoints x,y,visibility). There is no separate objectness channel. See the Hand Keypoints dataset docs and the general pose dataset overview.

      # out: (1, 300, 69)
      pred = out[0]                  # (300, 69)
      boxes_xyxy = pred[:, 0:4]      # x1,y1,x2,y2
      score = pred[:, 4]             # confidence score
      cls = pred[:, 5]               # class index (1 class)
      kpts = pred[:, 6:].reshape(-1, 21, 3)  # (300, 21, 3) -> x,y,vis

      If you tell me your export format (e.g. ONNX/TensorRT), you can drop the tensor names and a small snippet of the output here in this discussion so the team can see it.

  3. samasadiiGitHub(bearbeitet)

    I've noticed in this dataset almost all of the points are labeled as visible, even the ones that are not visible in the image.

  4. jugal-shethGitHub
    from ultralytics import YOLO
    
    # Load a model
    model = YOLO("yolo11n-pose.pt")  # load a pretrained model (recommended for training)
    
    # Train the model
    results = model.train(data="hand-keypoints.yaml", epochs=100, imgsz=640) 

    Error

    train: WARNING ⚠️ D:\projects\hand-track\datasets\dataset\train\images\IMG_00035498.jpg: ignoring corrupt image/label: cannot reshape array of size 67 into shape (2) train: WARNING ⚠️ D:\projects\hand-track\datasets\dataset\train\images\IMG_00035499.jpg: ignoring corrupt image/label: cannot reshape array of size 67 into shape (2) train: WARNING ⚠️ D:\projects\hand-track\datasets\dataset\train\images\IMG_00035500.jpg: ignoring corrupt image/label: cannot reshape array of size 67 into shape (2) . . . . . . . train: WARNING ⚠️ D:\projects\hand-track\datasets\dataset\train\images\IMG_00035501.jpg: ignoring corrupt image/label: cannot reshape array of size 67 into shape (2) train: WARNING ⚠️ D:\projects\hand-track\datasets\dataset\train\images\IMG_00035503.jpg: ignoring corrupt image/label: cannot reshape array of size 67 into shape (2) train: WARNING ⚠️ D:\projects\hand-track\datasets\dataset\train\images\IMG_00035504.jpg: ignoring corrupt image/label: cannot reshape array of size 67 into shape (2) train: WARNING ⚠️ D:\projects\hand-track\datasets\dataset\train\images\IMG_00035505.jpg: ignoring corrupt image/label: cannot reshape array of size 67 into shape (2) train: WARNING ⚠️ D:\projects\hand-track\datasets\dataset\train\images\IMG_00035507.jpg: ignoring corrupt image/label: cannot reshape array of size 67 into shape (2) train: WARNING ⚠️ D:\projects\hand-track\datasets\dataset\train\images\IMG_00035509.jpg: ignoring corrupt image/label: cannot reshape array of size 67 into shape (2) WARNING ⚠️ No images found in D:\projects\hand-track\datasets\dataset\train\labels.cache, training may not work correctly. See https://docs.ultralytics.com/datasets for dataset formattingg guidance. Traceback (most recent call last): File "d:\projects\hand-track\test.py", line 36, in model.train(data="config.yaml", epochs=100, imgsz=896, device=0) File "C:\Users\jugal\miniconda3\envs\hand\Lib\site-packages\ultralytics\engine\model.py", line 808, in train self.trainer.train() File "C:\Users\jugal\miniconda3\envs\hand\Lib\site-packages\ultralytics\engine\trainer.py", line 207, in train self._do_train(world_size) File "C:\Users\jugal\miniconda3\envs\hand\Lib\site-packages\ultralytics\engine\trainer.py", line 322, in _do_train self._setup_train(world_size) File "C:\Users\jugal\miniconda3\envs\hand\Lib\site-packages\ultralytics\engine\trainer.py", line 286, in _setup_train self.train_loader = self.get_dataloader(self.trainset, batch_size=batch_size, rank=LOCAL_RANK, mode="train") ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\Users\jugal\miniconda3\envs\hand\Lib\site-packages\ultralytics\models\yolo\detect\train.py", line 49, in get_dataloader dataset = self.build_dataset(dataset_path, mode, batch_size) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\Users\jugal\miniconda3\envs\hand\Lib\site-packages\ultralytics\models\yolo\detect\train.py", line 43, in build_dataset return build_yolo_dataset(self.args, img_path, batch, self.data, mode=mode, rect=mode == "val", stride=gs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\Users\jugal\miniconda3\envs\hand\Lib\site-packages\ultralytics\data\build.py", line 99, in build_yolo_dataset return dataset( ^^^^^^^^ File "C:\Users\jugal\miniconda3\envs\hand\Lib\site-packages\ultralytics\data\dataset.py", line 64, in init super().init(*args, **kwargs) File "C:\Users\jugal\miniconda3\envs\hand\Lib\site-packages\ultralytics\data\base.py", line 74, in init self.labels = self.get_labels() ^^^^^^^^^^^^^^^^^ File "C:\Users\jugal\miniconda3\envs\hand\Lib\site-packages\ultralytics\data\dataset.py", line 161, in get_labels len_cls, len_boxes, len_segments = (sum(x) for x in zip(*lengths)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ ValueError: not enough values to unpack (expected 3, got 0)

    1. Glenn JocherUltralytics(bearbeitet)

      @jugal-sheth hi there! It looks like your labels are incorrectly formatted for hand keypoints estimation. Each hand-keypoints label row needs 68 values: class, x, y, width, height (5 values) plus 21 keypoints with 3 values each (x, y, visibility = 63), but your rows appear to have 67. Please verify your label files match the format shown in the hand-keypoints documentation. Let us know if you need any other assistance!

    2. Glenn JocherUltralytics

      @jugal-sheth hi there! It seems your custom labels may not include the required bounding box coordinates alongside keypoints. The YOLO pose format requires each label to start with class, x_center, y_center, width, height (5 values) followed by 21 keypoints (63 values), totaling 68 values per label. Your 67-value labels suggest missing one element. Please verify your label format matches our pose estimation requirements exactly. Let us know if you need further clarification! 🚀

  5. berlintopGitHub

    I used yolov11n and yolov11x to train many models based on this dataset. The log accuracy of small batch and large epoch training is very good, but it is very poor during actual inference, much worse than mediaPipa, especially when rotating, the key points are very inaccurate. I am very confused about this, is there any trick to the training parameters?

    1. Glenn JocherUltralytics

      @737296 the issue you described could be due to overfitting or suboptimal augmentation and training practices. To address inaccuracies during inference, consider increasing data augmentation (e.g., rotation, scaling), fine-tuning learning rates or epochs, and ensuring sufficient diversity in the dataset. MediaPipe's robustness might stem from its architecture and preprocessing, so refining training parameters or exploring a different YOLO11 variant could help. Refer to the Hand Keypoints Dataset documentation for guidance on dataset setup and training. Let us know if further clarification is needed!

  6. chengziNaomiGitHub

    In yolov8-pose, what effect does the number of key points have on the detection results? Less key points, better results or more key points, better results? Now I want to identify the basket in the field, do I identify the four corner points on the top of the basket or the eight corner points on the basket?

    1. Glenn JocherUltralytics

      @chengziNaomi the number of keypoints can affect detection results based on your specific application. For identifying a basket, consider the complexity and precision needed. Using more keypoints might provide better detail, but it depends on your use case. For further guidance, you can explore the hand keypoints documentation at Ultralytics Docs.

    2. chengziNaomiGitHub

      Thank you for your reply. I have one more question. Why is the ratio of train and valid about 2:1 in this hand keypoints datasets? but in seg, the ratio is 10:1:1?Is it a special requirements from yolov8-pose?

    3. Glenn JocherUltralytics

      The dataset split ratios are often determined by the specific requirements of the task and dataset characteristics. For hand keypoints, a 2:1 ratio might balance training and validation needs effectively, while segmentation tasks might use a 10:1:1 ratio to ensure sufficient data for training, validation, and testing. These choices are not specific to YOLOv8-pose but are tailored to optimize model performance for each dataset type.

    4. chengziNaomiGitHub

      Thank you very much! I want to add negative samples to the datasets, what is the way for yolov8-pose to add negative samples?All I need to do is add the background image and blank txt text to train and val,right?

    5. Glenn JocherUltralytics

      To add negative samples in YOLOv8-pose, you can include background images with corresponding blank .txt files in your dataset. This approach helps the model learn to distinguish between relevant and irrelevant features.

  7. redsuningGitHub

    train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00001070.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0379] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00001796.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0006] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00001800.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0027] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00001801.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0049] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00001886.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0051 1.0038 1.0088 1.0114] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00001887.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0048 1.0062 1.0099 1.0121] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00001888.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1 1.0017] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00001994.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0059] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00001998.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0044] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00001999.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.007] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00002000.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0083] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00002001.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0146] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00002002.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0178] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00002003.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0178] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00002143.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0582 1.0216] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00002144.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0641 1.0333] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00002202.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0041 1.0052 1.008 1.0093 1.0108 1.0112 1.0116] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00002451.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.0121] train: WARNING ⚠️ G:\datasets\hand\hand-keypoints\train\images\IMG_00002454.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [ 1.023 1.0012] Many of the callout coordinates are beyond 1

    1. Glenn JocherUltralytics

      @redsuning the warnings indicate that some images have non-normalized or out-of-bounds coordinates. Please check your dataset annotations to ensure all keypoints are within the expected range. You might need to normalize the coordinates to fit within the [0, 1] range. For more details, you can refer to the dataset documentation.

    2. redsuningGitHub

      I know the reason for this warning, but the dataset I'm using is the default Hand-keypoints dataset you uploaded, and I mean your Hand-keypoints dataset labeling data needs to be modified.

      1
    3. Glenn JocherUltralytics

      Thank you for bringing this to our attention. We'll review the dataset annotations for any inconsistencies. If you have further questions, feel free to ask here.

    4. Glenn JocherUltralytics

      Thank you for pointing this out. We will investigate and address any labeling issues in the Hand Keypoints dataset as soon as possible. Meanwhile, you may exclude problematic samples or adjust the annotations for your use case to ensure optimal training results.